Advances in Computer Science for Engineering and Education IV (Lecture Notes on Data Engineering and Communications Technologies) [1st ed. 2021] 3030804712, 9783030804718

Citation preview

Lecture Notes on Data Engineering and Communications Technologies 83

Zhengbing Hu Sergey Petoukhov Ivan Dychka Matthew He   Editors

Advances in Computer Science for Engineering and Education IV

Lecture Notes on Data Engineering and Communications Technologies Volume 83

Series Editor Fatos Xhafa, Technical University of Catalonia, Barcelona, Spain

The aim of the book series is to present cutting edge engineering approaches to data technologies and communications. It will publish latest advances on the engineering task of building and deploying distributed, scalable and reliable data infrastructures and communication systems. The series will have a prominent applied focus on data technologies and communications with aim to promote the bridging from fundamental research on data science and networking to data engineering and communications that lead to industry products, business knowledge and standardisation. Indexed by SCOPUS, INSPEC, EI Compendex. All books published in the series are submitted for consideration in Web of Science.

More information about this series at http://www.springer.com/series/15362

Zhengbing Hu Sergey Petoukhov Ivan Dychka Matthew He •

•

•

Editors

Advances in Computer Science for Engineering and Education IV

123

Editors Zhengbing Hu School of Educational Information Technology Central China Normal University Wuhan, China Ivan Dychka Faculty of Applied Mathematics National Technical University of Ukraine “Igor Sikorsky Kiev Polytechnic Institute” Kiev, Ukraine

Sergey Petoukhov Mechanical Engineering Research Institute of the Russian Academy of Sciences Moscow, Russia Matthew He Halmos College of Natural Sciences and Oceanography Nova Southeastern University Plantation, FL, USA

ISSN 2367-4512 ISSN 2367-4520 (electronic) Lecture Notes on Data Engineering and Communications Technologies ISBN 978-3-030-80471-8 ISBN 978-3-030-80472-5 (eBook) https://doi.org/10.1007/978-3-030-80472-5 © The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Switzerland AG The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland

Contents

Computer Science for Manage of Natural and Engineering Processes Automatic Beam Aiming of the Laser Optical Reference System at the Center of Reflector to Improve the Accuracy and Reliability of Dynamic Positioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Serhii Zinchenko, Vladyslav Moiseienko, Oleh Tovstokoryi, Pavlo Nosov, and Ihor Popovych

3

Decision Support System in Sprinkler Irrigation Based on a Fractional Moisture Transport Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Vsevolod Bohaienko, Tetiana Matiash, and Anatolij Krucheniuk

15

Automated Pipeline for Training Dataset Creation from Unlabeled Audios for Automatic Speech Recognition . . . . . . . . . . . . . . . . . . . . . . . O. Romanovskyi, I. Iosifov, O. Iosifova, V. Sokolov, F. Kipchuk, and I. Sukaylo Estimation of Hurst Index and Traffic Simulation . . . . . . . . . . . . . . . . . Anatolii Pashko, Iryna Rozora, and Olga Syniavska Research of the Influence of Compromise Probability in Secure Based Traffic Engineering Model in SDN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Oleksandr Lemeshko, Zhengbing Hu, Anastasiia Shapovalova, Oleksandra Yeremenko, and Maryna Yevdokymenko

25

37

47

Model of Search and Analysis of Heterogeneous User Data to Improve the Web Projects Functioning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Solomiia Fedushko, Oleg Mastykash, Yuriy Syerov, and Anna Shilinh

56

Modeling the Dynamics of “Knowledge Potentials” of Agents Including the Stakeholder Requests . . . . . . . . . . . . . . . . . . . . . . . . . . . . Andrii Bomba, Taras Lechachenko, and Maria Nazaruk

75

v

vi

Contents

Crop Irrigation Demand Modelling Method Within the Conditions of Climate Change in the Dnipro-Donbas Canal . . . . . . . . . . . . . . . . . . Pavlo Kovalchuk, Viktoria Rozhko, Volodymyr Kovalchuk, Hanna Balykhina, and Olena Demchuk

89

The Method of Collision Risk Assessment Using Soft Safety Domains of Unmanned Vehicles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102 Volodymyr Sherstjuk, Maryna Zharikova, Igor Sokol, Irina Dorovskaja, Ruslan Levkivskiy, and Victor Gusev Content-Based Machine Learning Approach for Hardware Vulnerabilities Identification System . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Giorgi Iashvili, Zhadyra Avkurova, Maksim Iavich, Madina Bauyrzhan, Avtandil Gagnidze, and Sergiy Gnatyuk Diagnosis of Rail Circuits by Means of Fiber-Optic Cable . . . . . . . . . . . 127 N. Mgebrishvili, M. Iavich, Gr. Moiseev, N. Kvachadze, A. Fesenko, and S. Dorozhynskyy Method of Detecting a Fictitious Company on the Machine Learning Base . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 138 Hrystyna Lipyanina, Svitlana Sachenko, Taras Lendyuk, Vasyl Brych, Vasyl Yatskiv, and Oleksandr Osolinskiy Method for the Criticality Level Assessment for Crisis Situations with Parameters Fuzzification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 Andrii Gizun, Zhadyra Avkurova, Vladyslav Hriha, Anna Monashnenko, Nurbol Akatayev, and Marek Aleksander Analysis of Key Elements for Modern Contact Center Systems to Improve Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162 Bekmurza Aitchanov, Olimzhon Baimuratov, Muratbek Zhussupekov, and Tley Aitchanov End-to-End Scene Text Recognition Network with Adaptable Text Rectification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175 Yi Zhang, Zhiwen Li, Lei Guo, and Wenbi Rao Perfection of Computer Algorithms and Methods Forced Oscillations with Delays in Elastic and Damping Forces . . . . . . 187 Alishir A. Alifov Method of Static Optimization of the Process of Granulation of Mineral Fertilizers in the Fluidized Bed . . . . . . . . . . . . . . . . . . . . . . . 196 Bogdan Korniyenko and Lesya Ladieva Influence of the Software Development Project Context on the Requirements Elicitation Techniques Selection . . . . . . . . . . . . . . 208 Denys Gobov and Inna Huchenko

Contents

vii

AODV Protocol Optimization Software Method of Ad Hoc Network Routing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219 Liubov Oleshchenko and Kostiantyn Movchan Linear Planning Models Using Generalized Production Model of Leontief and Resource Constraints . . . . . . . . . . . . . . . . . . . . . . . . . . . 232 Alexander Pavlov, Iryna Mukha, Olena Gavrilenko, Liudmyla Rybachuk, and Kateryna Lishchuk Programming Style as an Artefact of a Software Artefacts Ecosystem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 244 Nikolay Sydorov Complex Model for Personal Data Management of Online Project Users . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256 Solomiia Fedushko, Oleg Mastykash, Yuriy Syerov, and Yaryna Kalambet Application of a Modified Logarithmic Rule for Forecasting . . . . . . . . . 270 Oleksandr Olefir and Orest Lastovsky A Survey on Kernelized Fuzzy Clustering Tools for Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281 Zhengbing Hu and Oleksii K. Tyshchenko Distributed Serverless Computing Orchestration Based on Finite Automaton . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290 Vadym Maslov and Andriy Petrashenko Matrix Stochastic Game with Q-learning for Multi-agent Systems . . . . . 304 Petro Kravets, Vasyl Lytvyn, Ihor Dobrotvor, Oleg Sachenko, Victoria Vysotska, and Anatoliy Sachenko Topology Synthesis Method Based on Excess De Bruijn and Dragonfly . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315 Heorhii Loutskii, Artem Volokyta, Pavlo Rehida, Artem Kaplunov, Bohdan Ivanishchev, Oleksandr Honcharenko, and Dmytro Korenko Method for Cyber Threats Detection and Identification in Modern Cloud Services . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 326 Zhengbing Hu, Sergiy Gnatyuk, Berik Akhmetov, Volodymyr Simakhin, Dinara Ospanova, and Nurbol Akatayev Nonparametric Change Point Detection Algorithms in the Monitoring Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347 Igor Prokopenko

viii

Contents

Computer Science for Education, Medicine and Biology Combinatorics of the Ford-Fulkerson Algorithm to Reduce the Risks of the COVID-19 Pandemic in International Tourism . . . . . . . . . . . . . . 363 Marharyta Sharko, Vira Fomishyna, Olha Liubchuk, Natalia Petrushenko, Tetiana Yakymchuk, and Liliia Chaika-Petehyrych Authentication System by Human Brainwaves Using Machine Learning and Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . 374 Z. B. Hu, V. Buriachok, M. TajDini, and V. Sokolov Approximate Training of Object Detection on Large-Scale Datasets . . . 389 Oleksandr Zarichkovyi and Iryna Mukha Framework for Developing a System for Monitoring Human Health in the Combined Action of Occupational Hazards Using Artificial Intelligence and IoT Technologies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 401 Oleksandra Yeremenko, Iryna Perova, Olena Litovchenko, and Nelia Miroshnychenko Chinese College Students’ L2 Motivational Self System in the Context of Artificial Intelligence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 411 Jie Ma and Pan Dong Encoding Sememe and Context Information into Sentence Representation for Implicit Sentiment Analysis . . . . . . . . . . . . . . . . . . . 423 Qizhi Qiu and Junan Qiu Course Resources and Teaching Methods of Logistics Management Major Under Emerging Engineering Education . . . . . . . . . . . . . . . . . . . 434 Yong Gu, Ju Chen, and Zhiping Liu Project-Oriented Course System Construction of Logistics Management Major . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444 Yue Liu, Yong Gu, and Zhiping Liu Analysis on the Venations, Hotspots and Trend of China’s Education Informatization Research in the Post-COVID-19 Era . . . . . . . . . . . . . . 454 Xiaofen Zhou, Yi Zhang, and Yanan Wang Author Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 467

Computer Science for Manage of Natural and Engineering Processes

Automatic Beam Aiming of the Laser Optical Reference System at the Center of Reflector to Improve the Accuracy and Reliability of Dynamic Positioning Serhii Zinchenko1 , Vladyslav Moiseienko1(B) , Oleh Tovstokoryi1 , Pavlo Nosov1 , and Ihor Popovych2 1 Kherson State Maritime Academy, Ushakova Avenue 20, Kherson 73000, Ukraine

[email protected] 2 Kherson State University, Universytetska st. 27, Kherson 73003, Ukraine

Abstract. This paper addresses the issues of automatic beam aiming of the laser optical reference system (LORS) at the center of reflector in conditions of strong pitching and rolling. Practical observations show that LORS does not receive a stable or high-quality reflection and sometimes even complete loss of the reflection is observed in conditions of strong waves. The authors have conducted full-scale experiments with the CyScan LORS, installed on a supply vessel “ESNAAD 225”, which confirmed instability and reflection loss under conditions of strong pitching and rolling. A method and algorithm were proposed for automatic aiming of the LORS beam into the center of reflector, which ensures a high reflection quality under conditions of strong pitching and rolling as well. The efficiency of automatic beam aiming under conditions of strong pitching and rolling was verified by mathematical modeling in MATLAB. The practical value of the obtained results lies in opportunity to improve the quality of the reflected signal and, therefore, to increase the accuracy and reliability of dynamic positioning in general, which is especially important when operating near hazardous objects such as oil and gas platforms. Keywords: Automatic beam aiming · Laser optical reference system · Offshore vessel · Reflector · Coordinate system

1 Introduction The dynamic position (DP) system is an automated onboard computer-controlled system that automatically maintains the required position and course of the vessel in order to reduce human error and risk of an accident during manual maneuvering [1–3]. DP system can work with both global (GPS) and LORS systems. GPS provides information about the geographical coordinates of the vessel’s location while LORS provides information about the vessel’s position relative to the landmark by measuring the bearing and distance to the landmark [4, 5]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 3–14, 2021. https://doi.org/10.1007/978-3-030-80472-5_1

4

S. Zinchenko et al.

The LORS occupies the second place in the list of installed reference systems, because it is characterized by accuracy, low cost, reliability, and ease of use, thus meeting the requirements of the International Maritime Organization, International Association of Marine Contractor and Det Norske Veritas Germanischer Lloyd [6–8]. The laser optical reference system is a measuring system and integral part of the DP system. It consists of a laser unit, integrated controller, and intelligent software [9, 10]. The laser unit rotates or scans in a horizontal plane and emits high-frequency pulses. The pulses are reflected from the reflectors attached to moving or stationary objects (platforms or vessels) and returned back to the laser unit. The distance to the object is measured by the difference in the time of emission and return of pulses [11]. Laser optical reference system, as any other positioning system, has its own risks of losing position. One of the reasons for signal loss is shading of the reflector by objects or contamination of the lens. Such problems are easily solved by correct installing the reflector (in the line of sight), cleaning the lens of dirt, etc. [12, 13]. However, deviation of the optical axis outside the reflector, especially in conditions of strong pitching and rolling, due to the limited beam width (12–18°) in the vertical plane [14–16] appears to be more significant reason of losing position, which leads to the weakening and even disappearance of the signal. The risk of collision is associated with the human factor as well. The influence of the human factor on safety has been considered by authors, in particular [17, 18]. Reducing the influence of the human factor can also be achieved through the creation of automated [19, 20] or automatic [21] ship traffic control systems. The works [22–26] consider the issues of creating such systems for other modes of transport too.

2 Review of Literature Sources Focused on Improving the Accuracy and Reliability of Dynamic Positioning The issues of increasing the accuracy and reliability of automatic control systems, including DP systems, were considered in many research works, for example, [27–31]. Thus, the article [27] explored a new method for determining location in visible light using a dynamic positioning and tracking algorithm based on optical flow detection and Bayesian prediction. Experiments showed that the proposed dynamic positioning and tracking algorithm could provide a high positioning accuracy and reduce the risk of accident. The article [28] studied the accuracy of smartphones when maneuvering a vessel in the dynamic positioning system. The researcher’s results showed that in this case the positioning accuracy of 2–3 cm was provided with a probability of 95%. The article [29] considered a reliable nonlinear control law for the dynamic positioning system using a perturbation observer, an auxiliary dynamic system, and a dynamic surface control method. The developed law of reliable nonlinear DP control proved the ability to maintain the position of the vessel and to reduce risk of collision, while guaranteeing a uniform limitation of all signals in a closed DP control system. The article [30] studied a nonlinear adaptive fuzzy feedback controller for dynamic positioning system in order to reduce the risk of losing position. The proposed nonlinear adaptive feedback controller with fuzzy output showed improvement of stability of positioning under various environment conditions.

Automatic Beam Aiming of the Laser Optical Reference System

5

The user manual [31] describes three DP systems: Navis Engineering OY, Marine Technologies LLC and Rolls Royce Marine. Each of these DP systems works with LORS. At the same time, before starting operations with LORS, the angle of elevation of the LORS beam above the horizon is manually set up in each of these systems, so that the beam hits the reflector while the laser direction unit is passing the reflector during circular or sector scanning. As follows from the above review, the solutions do not consider the possibility of increasing the accuracy and reliability of the DP system by automatically aiming the LORS beam at the center of the reflector. The closest solutions [31] involve only the initial manual setting of the beam elevation, which is insufficient during strong pitching or rolling. Therefore, the development of method, algorithm and software for the system of automatic aiming of the beam at the center of the reflector, considering the angular deviations of the vessel from the initial position during rolling and pitching, is an urgent scientific and technical problem. The object of research is the processes of automatic aiming of the LORS beam at the center of the reflector under conditions of strong rolling and pitching. The subject of research is method, algorithm and software for the control system of aiming in conditions of strong pitching and rolling. The purpose of the research is to improve the accuracy and reliability of the dynamic positioning system in conditions of strong pitching and rolling. This is achieved by automatically aiming the LORS beam at the center of the reflector in conditions of strong pitching and rolling, namely: LORS measurements of bearing Bm (n) and distance Dm (n) to the reflector at the moment the beam is passing the reflector; according to the measured bearing Bm (n), distance Dm (n) and the elevation h2 − h1 of the reflector center above the laser tube, the elevation angle θ ∗ is determined; measured roll ϕm (n) and pitch ϑm (n) angles of the vessel, together with the elevation angle θ ∗ , are used to calculate the required optical axis position angle θm0 (n) in vertical plane; the current position of the optical axis θm (n) is brought to the required position θm0 (n).

3 Method and Algorithms for Automatic Aiming of the LORS In the process of our research, were applied the following methods: analysis and synthesis, methods of theory of automatic control systems, linear algebra methods, matrix calculus and transformation of coordinate systems, methods of numerical integration, mathematical modeling methods, and full-scale experiment methods. The linked coordinate system (LCS) XL YL ZL associated with the platform is located in the center of rotation of the vessel, the axis OXL lies in the plane of the local horizon and is directed towards the reflector, the axis OYL lies in the plane of the local horizon being perpendicular to the axis OXL and directed to the right. The axis OZL complements the system to the “right”. The base coordinate system (BCS) XB YB ZB , associated with the vessel, is located in the center of rotation of the vessel, the axis OXB lies in the diametrical plane and is directed towards the stern of the vessel, the axis OYB is perpendicular to the axis OXB and directed to the port side, the axis OZB complements the system to the “right”.

6

S. Zinchenko et al.

The measuring coordinate system (MCS) XM YM ZM is located in the center of rotation of the vessel, the axis OXM is directed along the optical axis, the axis OYM is perpendicular to the axis OXM and directed to the right, the axis OZM complements the system to the “right”. Figure 1 shows the coordinate systems used in the calculations.

Fig. 1. Coordinate systems

The transition matrix A between LCS and BCS for the sequence of turns Z − Y − X is presented in Table 1, where ϕ, ψ, ϑ are the angles of roll, yaw and pitch, respectively. Table 1. Transition matrix A between LCS and BCS A

YL

ZL

XB cos ϑ cos ψ

cos ϑ sin ψ

− sin ϑ

YB

sin ϕ sin ϑ cos ψ−

sin ϕ sin ϑ sin ψ+

cos ϕ sin ψ

cos ϕ cos ψ

sin ϕ sin ψ+

cos ϕ sin ϑ sin ψ−

cos ϕ sin ϑ cos ψ

sin ϕ cos ψ

ZB

XL

sin ϕ cos ϑ

cos ϕ cos ϑ

The transition matrix B between the BCS and the MCS for the sequence of turns Z − Y is presented in Table 2, where ψm , θm are the angles that determine the position of the optical axis in a BCS. The unit vector, defining the required optical axis direction to the center of the reflector in LCS, is (see Fig. 1). eLCS = (cos θ ∗ , 0, − sin θ ∗ ), θ ∗ = arcsin(

h2 − h1 ), Dm

Automatic Beam Aiming of the Laser Optical Reference System

7

Table 2. Transition matrix B between BCS and MCS B

XB

YB

ZB

XM

cos θm cos ψm cos θm sin ψm − sin θm

YM

− sin ψm

cos ψm

0

ZM

sin θm cos ψm

sin θm sin ψm

cos θm

where h1 , h2 are the laser and reflector heights above sea level, Dm is the measured distance to the center of the reflector. The unit vector, defining the required optical axis direction to the center of the reflector in BCS, taking into account Table 1, has the form eyBCS ezBCS

exBCS = cos θ ∗ cos ϑ cos ψ + sin θ ∗ sin ϑ, ∗ cos θ (sin ϕ sin ϑ cos ψ − cos ϕ sin ψ) − sin θ ∗ (sin ϕ cos ϑ),

= = cos θ ∗ (sin ϕ sin ψ + cos ϕ sin ϑ cos ψ) − sin θ ∗ (cos ϕ cos ϑ).

The unit vector, defining the optical axis direction in BCS, taking into account Table 2, has the form eBCS = (cos θm cos ψm , cos θm sin ψm , − sin θm ). Conditions, that determine the direction of the optical axis to the center of the reflector are written down cos θm0 cos ψm0 = cos θ ∗ cos ϑ cos ψ + sin θ ∗ sin ϑ, cos θm0 sin ψm0 = cos θ ∗ (sin ϕ sin ϑ cos ψ − cos ϕ sin ψ) − sin θ ∗ (sin ϕ cos ϑ), − sin θm0 = cos θ ∗ (sin ϕ sin ψ + cos ϕ sin ϑ cos ψ) − sin θ ∗ (cos ϕ cos ϑ).

(1)

3.1 Aiming the LORS Beam in the Laser Rotation or Scan Mode In the case under consideration, the bearing to the reflector is measured only at the moment when the beam is passing the reflector, the rest of the time the bearing is predicted using the measured yaw rate. 0 (n) + ω (n)T ; B(n)  = B (n), Bm (n) = Bm zm m 0 (n); B (n) = B0 (n), Bm (n) = Bm m m 0 (n) is where Bm (n) is the bearing prediction to the reflector at the n - calculation step, Bm the measured bearing to the reflector at the n - calculation step, ωzm (n) is the measured yaw rate at the n - calculation step, T is the information processing cycle in the onboard controller. From the last equation of system (1), taking into account for the adopted coordinate systems, that Bm (n) = −ψ, ϕm (n) = ϕ, ϑm (n) = ϑ, is determined the following

θm0 (n) = arcsin (sin θ ∗ cos ϕm (n) cos ϑm (n) + cos θ ∗ (sin ϕm (n) sin Bm (n)− cos ϕm (n) sin ϑm (n) cos Bm (n))),

(2)

8

S. Zinchenko et al.

where θm0 (n) is the required optical axis deviation angle in vertical plane at the n calculation step, ϕm (n) is the measured roll angle of the vessel at the n - calculation step, ϑm (n) is the measured pitch angle of the vessel at the n - calculation step. Equations (2) determine the required angle θm0 (n) of the optical axis deviation in vertical plane at the n - calculation step, at which the optical axis will be directed to the center of the reflector for any measured values of the roll ϕm (n) and pitch ϑm (n) angles of the vessel at the n - calculation step. Reduction of the beam to the required position is performed according to the difference scheme, which is a discrete analogue of the aperiodic link θm (n) = θm (n − 1) +

T 0 (θ (n) − θm (n)), T m

(3)

where T is the time constant of aperiodic link. 3.2 Aiming the LORS Beam in the Follow Reflector Mode In the case under consideration, the laser module is constantly oriented to the reflector by changing the direction of the beam in both the vertical and horizontal planes. From the last equation of system (1), taking into account, that Bm (n) = −ψ, ϕm (n) = ϕ, ϑm (n) = ϑ, is determined the following θm0 (n) = arcsin (sin θ ∗ cos ϕm (n) cos ϑm (n) + cos θ ∗ (sin ϕm (n) sin Bm (n)− cos ϕm (n) sin ϑm (n) cos Bm (n))),

(4)

From the first and second equations of system (1), is determined the following sin θm (n) cos Bm (n)+cos ϕm (n) sin Bm (n) ψm0 (n) = arct ( sin ϕm (n) cos θm (n) cos Bm (n)+tgθ ∗ sin θm (n) ∗

sin ϕm (n) cos θm (n) − cos θmtgθ (n) cos Bm (n)+tgθ ∗ sin θm (n) ).

(5)

Equations (4), (5) determine the required angles θm0 (n), ψm0 (n) of the optical axis deviation in vertical and horizontal plane at the n - calculation step, at which the optical axis will be directed to the center of the reflector for any measured values of the roll ϕm (n) and pitch ϑm (n) angle of the vessel at the n - calculation step. The reduction of the beam to the required position is performed according to the difference scheme, which is a discrete analogue of the aperiodic link θm (n) = θm (n − 1) +

T 0 (θ (n) − θm (n)), T m

(6)

ψm (n) = ψm (n − 1) +

T 0 (ψm (n) − ψm (n)). T

(7)

Automatic Beam Aiming of the Laser Optical Reference System

9

4 Experiment 4.1 Full-Scale Experiments with Laser Optical Reference System CyScan on the Vessel “ESNAAD 225” The authors conducted an experiment with LORS CyScan, installed on the platform supply vessel “ESNAAD 225” and reflector mounted on a Jack-Up type platform. Figure 2 shows the LORS CyScan screen during experiments with pitching and rolling. The LORS CyScan screen are displays the position of the reflector relative to the vessel, the bearing and the distance to the target, the quality of the received data, the brightness of the reflecting rays, and the angle of the beam relative to the horizon. Experiment with Pitching and Rolling. Distance to the reflector is 68.8 m, bearing to the reflector is 200.9°, tracking mode is a single reflecting tube, beam tilt is in auto mode. Figure 2a shows the LORS CyScan screen with pitching (pitching amplitude is 6°–9°), pitching period is 3–4 s). Figure 2b shows the LORS CyScan screen with rolling (rolling amplitude is 11°–14°), rolling period is 1.5–2 s. The below figure shows, the LORS CyScan reflection during pitching is unstable, the quality and brightness of reflection is low. The LORS CyScan reflection during rolling is also unstable, the quality and brightness of the reflection is minimal, a loss of the reflection from the reflector tube occur.

Fig. 2. Experiment with pitching and rolling

Since the system can be installed on any type of the vessel (DSV, PSV, or Shuttle tanker) [32, 33], an unstable reflection or its complete loss can lead to disruption of the reference positioning system and failure as a whole [34, 35], which is associated with severe risks: environmental pollution, accidents on the oil and gas platform, and even human casualties.

10

S. Zinchenko et al.

4.2 Simulation of Automatic Beam Aiming at Reflector in MATLAB The operability and efficiency of the proposed method and algorithms for automatic aiming of the beam at the center of the reflector was tested by mathematical modeling in the MATLAB environment in the follow reflector mode. Figure 3 and 4 demonstrate the results of mathematical modeling of the processes of dynamic positioning of the vessel in the form of graphs of changes in time longitudinal speed Vx , lateral speed Vy , longitudinal displacement Xg , lateral displacement Yg , roll rate ωx (wx), roll angle ϕ(fi), pitch rate ωy (wy), pitch angle ϑ(tet), yaw rate ωz (wz), yaw angle ψ(psi), bearing to the platform Pm , distance to the platform Dm , angles, that determine the current position of the optical axis LORS in BCS θm (tetCyScan), ψm (psiCyScan) and optical axis deviations from the direction to the center of reflector in vertical plane (dtet) and horizontal plane (dpsi). Figure 3 provides simulation results for pitching parameters (amplitude and frequency), which coincide with the corresponding parameters in a fullscale experiment. Furthermore, in order to create more severe conditions compared to the full-scale experiment, the initial values of the longitudinal (Vx ), lateral (Vy ) and angular (wz) ship speeds are set.

Fig. 3. Simulation results of the automatic beam aiming in conditions of strong pitching

Figure 4 present a simulation results of the automatic beam aiming at the center of the reflector under conditions of strong rolling. The results are given for the parameters of rolling (amplitude and frequency), which coincide with the corresponding parameters in a full-scale experiment. Furthermore, in order to create more severe conditions compared to the full-scale experiment, the initial values of the longitudinal (Vx ), lateral (Vy ) and angular (wz) ship speeds are set as well.

Automatic Beam Aiming of the Laser Optical Reference System

11

Fig. 4. Simulation results of the automatic beam aiming in conditions of strong rolling

The above results prove that the DP system ensures that the given initial speeds and the resulting displacements (Xg , Yg , psi) are processed to zero in the presence of strong oscillations in the roll channel (ωx (wx), ϕ(fi)) and trim (ωy (wy), ϑ(tet)) from pitching or rolling. The results of comparing the quality of the reflected signal Q for the known solution (CyScan) and the proposed (automatic beam aiming) are shown in Table 3. Table 3. Results of comparing the quality of the reflected signal Type of vessel motions Pitching

Vessel motions parameters A [dgr] 6–9

CyScan Q [%]

Automatic aiming

T [sec.]

dtet, dpsi

Q [%]

3–4

|dtet| < 2◦ ,

100

|dtet| < 2◦ , |dpsi| < 0, 25◦

100

30

|dpsi| < 0, 25◦ Rolling

11–14

1,5–2

0

The above results indicate that, for the amplitudes A and periods T of pitching and rolling indicated in the table, the quality of the reflected signal Q of the known solution (CyScan) is, respectively, 30% and 0% (signal loss) (see the data of the fullscale experiment in Fig. 2). For the same pitching and rolling parameters, the quality Q of the reflected signal with automatic beam aiming in both cases is 100%, since the deviation of the beam from the center of the reflector in the vertical plane |dtet| < 2◦ is 6 times less than the beam width (12°) in the vertical plane, and the deviation of the beam in the horizontal plane |dpsi| < 0, 25◦ is commensurate with the width of the beam in the horizontal plane (0.13°). This means that the accuracy obtained by automatic beam aiming allows using not only Laser Rotation or Scan Mode, in which the beam position can be refined with a period of rotation or scanning of the laser unit, but even the Follow

12

S. Zinchenko et al.

Reflector Mode, in which the beam position can be refined on each clock cycle of the onboard controller.

5 Conclusion The proposed method, algorithm and software for the system of automatic aiming of the LORS beam at the center of the reflector allow to significantly improve the accuracy and reliability of the dynamic positioning system as a whole in conditions of strong pitching and rolling. The known solutions implemented in the DP systems of Navis Engineering OY, Marine Technologies LL and Rolls Royce provide only the initial manual setting of the LORS optical axis elevation angle, which is not enough under conditions of strong pitching and rolling. As shown by the results of the full-scale experiment conducted on the vessel ESNAAD 225 with LORS CyScan, during strong pitching or rolling, the quality of the signal reflected from the reflector deteriorates significantly, and in some cases, it may even be lost. Such work of LORS is unacceptable when performing dynamic positioning or maneuvering operations near oil and gas platforms. Deterioration of the reflected signal quality or loss occur because of the excessive deviation of the signal from the center of the reflector in case of strong waves. The scientific novelty of obtained results lies in theoretical justification of design features of the original system of automatic aiming of the LORS beam at the center of the reflector, consisting in constant, on each clock cycle of the onboard controller measuring the angular displacements of the vessel from the basic position. Subsequent consideration of these deviations is used to calculate the direction of the optical axis to the center of the reflector. Bringing the current position of the optical axis to the calculated position is performed by automatically changing the elevation angle of the optical axis. The obtained results allow to significantly improve the quality of the reflected signal, to exclude signal loss under conditions of strong pitching and rolling. The practical value of the results obtained lies in the development and introduction into industrial production of the original system of automatic aiming of the LORS beam at the center of the reflector as well as the regulatory documentation for it being able to improve the quality of the reflected signal. Therefore, accuracy and reliability of the dynamic positioning system in general under conditions of strong pitching and rolling can be achieved. As a consequence, different kinds of risks can be reduced when performing operations near hazardous objects, including an oil and gas platform, which as a whole proves the practical value of the results obtained. Further research will be related to improving the quality of automatic aiming by taking into account the statistical characteristics of wave.

References 1. IMCA Marine Division: Guidelines for the design and operation of dynamically positioned vessels. The International Marine Contractors Association (2019) 2. Bray, D., Daniels, J., Fiander, G., Foster, D.: DP Operator’s Handbook The Nautical Institute (2020) 3. Requirements for vessels and units with dynamic positioning (DP) system. Polish Register of Shipping (2019)

Automatic Beam Aiming of the Laser Optical Reference System

13

4. The Safe Operation of Dynamically Positioned Offshore Supply Vessels. The International Marine Contractors Association. IMCA Marine Division (2018) 5. Review of the Use of the Fan Beam Laser System for Dynamic Positioning. The International Marine Contractors Association. IMCA Marine Division (2017) 6. Guidelines for vessels and units with dynamic positioning (DP) systems. IMO. MSC.1/Circ.1580 (2017) 7. Dynamic positioning vessel design philosophy guidelines. Det Norske Veritas Germanischer Lloyd (2015) 8. Rules for Classification and Construction Ship Technology. Det Norske Veritas Germanischer Lloyd (2016) 9. Meyers, J., Hoburg, A.: A capstone project on robust dynamic positioning and data acquisition systems. In: Proceedings of 2015 ASEE Northeast Section Conference (2015) 10. Palmieri, S., Kline, D., Emami, T., Reza, A., Swaszek, P.: Multithreaded sensor analysis for a redundant dynamic positioning 2 project. In: Proceedings of the ASEE NE 2016 Conference (2016) 11. Erdman, J.: Sea Vipers computer vision and inertial position reference sensor system (CVIPRSS). Louisiana State University (2015) 12. McKnight, D., Miles, R.: Impact of reduced visibility conditions on laser-based DP sensors. In: Dynamic Positioning Conference (2014) 13. Dong, Y., Vinnem, J.E., Utne, I.B.: Improving safety of DP operations: learning from accidents and incidents during offshore loading operations. EURO J. Decis. Process. 5(1–4), 5–40 (2017). https://doi.org/10.1007/s40070-017-0072-1 14. Renishaw Apply Innovation. Fanbeam 5 Operator’s manual (2017) 15. Sigmund, S., Moctar, O.: Numerical and experimental investigation of added resistance of different ship types in short and long waves. Ocean Eng. 147, 51–67 (2018). https://doi.org/ 10.1016/j.oceaneng.2017.10.010 16. Xu, S., Wang, X., Wang, L.: Mitigating roll–pitch motion by a novel controller in dynamic positioning system for marine vessels. J. Ships Offshore Struct. 12(8), 1136–1144 (2017). https://doi.org/10.1080/17445302.2017.1316905 17. Popovych, I., et al.: Research on the types of pre-game expectations in the athletes of sports games. J. Phys. Educ. Sport 20(1), 43–52 (2020). https://doi.org/10.7752/jpes.2020.01006 18. Popovych, I., Blynova, O., Aleksieieva, M., Nosov, P., Zavatska, N., Smyrnova, O.: Research of relationship between the social expectations and professional training of lyceum students studying in the field of shipbuilding. Revista ESPACIOS 40(33), 21 (2019) 19. Nosov, P., Ben, A., Zinchenko, S., Popovych, I., Mateichuk, V., Nosova, H.: Formal approaches to identify cadet fatique factors by means of marine navigation simulators. In: CEUR Workshop Proceedings, vol. 2732, pp. 823–838 (2020) 20. Nosov, P., Zinchenko, S., Popovych, I., Safonov, M., Palamarchuk, I., Blah, V.: Decision support during the vessel control at the time of negative manifestation of human factor. In: CEUR Workshop Proceedings, vol. 2608, pp. 12–26 (2020) 21. Zinchenko, S., Tovstokoryi, O., Nosov, P., Popovych, I., Kobets, V., Abramov, G.: Mathematical support of the vessel information and risk control systems. In: CEUR Workshop Proceedings, vol. 2805, pp. 335–354 (2020) 22. Fallooji, N., Abbod, M.: Helicopter control using fuzzy logic and Narma-L2 techniques. Int. J. Intell. Syst. Appl. 5(12), 1–14 (2020). https://doi.org/10.5815/ijisa.2020.05.01 23. Youssef, F., Houda, B.: Comparative study of end-to-end deep learning methods for selfdriving car. Int. J. Intell. Syst. Appl. 5(12), 15–27 (2020). https://doi.org/10.5815/ijisa.2020. 05.02 24. Hu, Z., Khokhlachova, Y., Sydorenko, V., Opirskyy, I.: Method for optimization of information security systems behavior under conditions of influences. Int. J. Intell. Syst. Appl. 12(9), 46–58 (2017). https://doi.org/10.5815/ijisa.2017.12.05

14

S. Zinchenko et al.

25. Rasim, M., Alguliyev, R., Aliguliyev, Y., Imamverdiyev, L.: Sukhostat: an anomaly detection based on optimization. Int. J. Intell. Syst. Appl. 12(9), 87–96 (2017). https://doi.org/10.5815/ ijisa.2017.12.08 26. Krishan, G., Singh, V.: Motion control of five bar linkage manipulator using conventional controllers under uncertain conditions. Int. J. Intell. Syst. Appl. 5(8), 34–40 (2016). https:// doi.org/10.5815/ijisa.2016.05.05 27. Weipeng, G., Xin, C., Mouxiao, H., Zixuan, L., Yuxiang, W., Yingcong, C.: High-speed robust dynamic positioning and tracking method based on visual visible light communication using optical flow detection and Bayesian forecast. IEEE Photonics J. 10(3) (2018). https://doi.org/ 10.1109/JPHOT.2018.2841979 28. Specht, C., Dabrowski, P.S., Pawelski, J., Specht, M., Szot, T.: Comparative analysis of positioning accuracy of GNSS receivers of Samsung Galaxy smartphones in marine dynamic measurements. Adv. Space Res. 63(9), 3018–3028 (2019). https://doi.org/10.1016/j.asr.2018. 05.019 29. Jialu, D., Xin, H., Miroslav, K., Yuqing, S.: Robust dynamic positioning of ships with disturbances under input saturation. Automatica 73, 207–214 (2016). https://doi.org/10.1016/j.aut omatica.2016.06.020 30. Xiaogong, L., Jun, N., Yuzhao, J., Kun, L., Heng, L.: Nonlinear adaptive fuzzy outputfeedback controller design for dynamic positioning system of ships. Ocean Eng. 158, 186–195 (2018). https://doi.org/10.1016/j.oceaneng.2018.03.086 31. Navi – Trainer 5000. Transas offshore simulator. Manual. Transas MIP Ltd. (2012) 32. Chae, C., Jung, Y.: A study on FSA application to PRS for safe operation of dynamic positioning. Vessel J. Navig. Port Res. 41(5), 287–296 (2017). https://doi.org/10.5394/KINPR. 2017.41.5.287 33. Pil, I.: Causes of Dynamic Positioning System Failures and Their Effect on DP Vessel Station Keeping. Tallinn University of Technology Estonian Maritime Academy (2018) 34. Olubitan, O., Loughney, S., Wang, J., Bell, R.: An investigation and statistical analysis into the incidents and failures associated with dynamic positioning systems. In: Safety and Reliability – Safe Societies in a Changing World (2018). https://doi.org/10.1201/9781351174664-10 35. Chong, C., Chul, Y.: A study on FSA application to PRS for safe operation of dynamic positioning vessel. J. Navig. Port Res. 41(5), 287–296 (2017). https://doi.org/10.5394/KINPR. 2017.41.5.287

Decision Support System in Sprinkler Irrigation Based on a Fractional Moisture Transport Model Vsevolod Bohaienko1(B) , Tetiana Matiash2 , and Anatolij Krucheniuk1 1 VM Glushkov Institute of Cybernetics of NAS of Ukraine, Kyiv, Ukraine

[email protected] 2 Institute of Water Problems and Land Reclamation of NAASU, Kyiv, Ukraine

Abstract. The paper presents a novel algorithm for decision support in sprinkler irrigation and its software implementation. The proposed algorithm is based on the modeling of moisture transport using a fractional differential generalization of the Richards equation stated in terms of water head and on the usage of particle swarm optimization for model calibration. The paper also describes the algorithm’s implementation that contains a field installed hardware part that monitors soil and surface air condition, and an analytical software part that stores and processes monitoring data in order to provide recommendations on irrigation schedules and rates. The proposed technique allows to increase the simulation accuracy up to 7% while modeling 3 months vegetative period enabling essential increase in the recommendation adaptability to the changing vegetation conditions. Keywords: Decision support algorithm · Irrigation management · Moisture transport modeling · Richards equation · Fractional differential model

1 Introduction Climate change causes a continuous deterioration of natural water supply conditions during crop cultivation and urges a growing need for effective irrigation. As the management level increasingly determines the efficiency and environmental safety of irrigation, it is important to improve and develop new irrigation management technologies among which decision support systems (DSSs) play the central role. DSSs in irrigation are the tools aimed first and foremost at defining the irrigation schedules and rates. These systems can be divided into two classes. The first one is aimed at performing automated irrigation using sensor readings only [1, 2]. The second one models and predicts the state of soil moisture giving the valuable information to the farmers especially in case when sprinkling irrigation, which cannot be started immediately, is used. Mainly, such DSSs [3] are based on balance models of soil moisture content dynamics. The simplest model is the single layer balance model that takes evapotranspiration estimates and water income volumes as inputs. Further developments generalize the single layer model to multilayered one and to differential hydrological ones in order to take the process of water deep percolation and other processes in the system “soil-plant-atmosphere” into account. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 15–24, 2021. https://doi.org/10.1007/978-3-030-80472-5_2

16

V. Bohaienko et al.

Differential models used in DSSs in irrigation describe moisture transport in soil. They are mostly based on the Richards equation stated in terms of moisture content or water head. This equation considers the soil with the properties independent of scale. Since soils can be considered as fractal-structured media in case of complex hydrogeological conditions [4], the use of the integer-order Richards equation can lead to substantial inaccuracies while predicting moisture dynamics that can further cause yield loses. Recalling that numerous results were recently obtained in the field of mathematical modeling of mass transfer processes in fractal-structured porous media using the formalism of fractional-order differential operators [5, 6], the aim of our research is to construct a novel algorithm and a software on its base for predicting irrigation schedules and rates performing moisture transfer modeling using some fractional differential Richards equation.

2 Literature Review As a widely used example of DSS in irrigation [3, 7] based on a single layer moisture content model, the AquaCrop [8] system implemented according to the FAO-33 [9] and FAO-66 [10] documents can be mentioned. One of the main factors influencing the accuracy in the mentioned and other DSSs is evapotranspiration estimation performed in AquaCrop using the Penman-Monteith [11] method. Among the evapotranspiration estimation methods used in Ukraine the methods of Shtoyko [12] and Ivanov [13] can also be singled out. Since effective irrigation management requires to the fullest extent complete consideration of all processes in the system “soil-plant-atmosphere”, DSSs in irrigation often take into account such issues as nitrogen cycles [14, 15], surface and groundwater pollution [16], risk assessment and economic modeling [17]. Among the known systems CropWat [18], IrriSatSMS [19], DAISY [20], CropIrri [21], APSIM [22], CropSyst [23] can be mentioned. Only some of them, in particular DAISY [20] and APSIM [22], use classical hydrological models to predict the state of soil moisture thus making topical further evolution of DSSs in irrigated agriculture enlarging the modeling base and increasing the prediction accuracy.

3 Decision Support Algorithm The accurate prediction of soil moisture state requires a nexus of moisture transport and evapotranspiration models with algorithms for their calibration that forms the core of a decision support system. Thus, the methodology used in our research is to combine a fractional differential moisture transport equation with metaheuristic approach for determining its non-measurable parameters including a factor on which evapotranspiration estimates are multiplied. Such approach is from the best of our knowledge novel in the field of decision support in irrigated agriculture. We propose to use a one-dimensional differential Richards equation stated in terms of water head in a generalized fractional differential formulation as fitting of its additional

Decision Support System in Sprinkler Irrigation

17

parameters allows model’s adaptation to specific field conditions and minimizes the influence of input data inaccuracies. The model has the following form [7]:   ∂H (β) Dt H = C −1 (H ) Dz(α) (k(H ) ) − S , 0 ≤ z ≤ L, t ≥ 0 (1) ∂z where H is the water head (m), C −1 (H ) = ∂H ∂θ , θ is the volumetric moisture content of soil (%), H (θ ) is the water retention curve, k(H ) is the hydraulic conductivity, S is the moisture extraction function that models the interaction between the root system (β) of plants and soil, Dt is the Caputo-Gerasimov fractional derivative with respect to t  (β) (α) 1 the time variable in the form Dt H (t, z) = Γ (1−β) H (t, z)(t − τ )−β d τ , Dz is the 0

similar derivative with respect to the space variable, 0 < α, β ≤ 1. For the water head function H at the lower boundary z = L of the solution domain, in the case of the groundwater horizon on a confining bed, the impermeability condition ∂H ∂z = 0 is set. In the absence of groundwater and a confining bed, we set the first-order condition H = HL where HL is the given function. At the upper boundary z = 0, when soil is in saturated state and water accumulates on its surface, the first-order boundary condition H = HU is set (HU is the given function.) In other situations, the following second-order boundary condition is set: k

∂H = Qe − Qp − Qi ∂z

where Qe , Qp , Qi are the fluxes caused by evaporation, precipitation, and irrigation. The moisture extraction function S uses the plant’s root system density function to distribute volume of moisture transpirated by plants within the root layer of soil. Evaporation and transpiration fluxes are determined on the base of leaf area index from actual evapotranspiration estimate [25]. The latter is estimated using the formulas of Ivanov [13] and Shtoyko [12] on the base of measurements of surface air temperature and humidity. Leaf area index is assessed using remote sensing data. To take into account evapotranspiration’s estimation errors, its value is multiplied by a coefficient that can be further chosen in the process of model’s fitting to experimental data. Precipitation and irrigation fluxes are also multiplied by a coefficient that models measurement errors. Soil’s water retention curves and the dependencies between hydraulic conductivity and water head, which are the input parameters of the model, are specified per soil layer according to the van Genuchten’s model [24]. Its coefficients are fitted to conform to experimental soil survey data [25]. In the absence of the latter, known approximating methods implemented in the Rosetta software [26] are used with particle size distribution and soil density given as inputs. For the most accurate modeling, the parameters of plants’ root system development—its depth and roots density distribution—should also be evaluated. We solve the initial-boundary value problem for the Eq. (1) with the first-order initial condition by a finite difference method, similarly to the described in [27, 28].

18

V. Bohaienko et al.

The model (1) is calibrated [25] by selecting the values of its parameters (fractional derivatives orders, evapotranspiration multiplier value, precipitation and irrigation fluxes multiplier value) to obtain the conformance to the measured water head dynamic. The particle swarm optimization (PSO) algorithm is used for this purpose [29] as such an inverse problem does not have an analytic solution and its parameters are real numbers. Rigorous description of the used variety of the PSO algorithm is given in [30]. The set of model’s parameters that are fitted can be extended if some of soil’s hydrophysical properties are unknown. From the other hand, a large number of parameters can lead to overfitting. As the Eq. (1) is the generalization of the classical Richards equation, we can, beside its fractional form, also use the classic form within the same algorithmic base and software to have another tool that is less vulnerable to overfitting. To secure continuous decision support in the case of input data sources failure, we also use a simple water balance model (e.g. [31]) that does not take water deep percolation into account. Having the models of moisture transport we can perform hydrological calculations to predict soil moisture state. On its base we schedule the next irrigation when an averaged suction pressure value in the root layer of soil reduces to a given pre-irrigation threshold. Irrigation continues until the upper limit of the maintainable range is reached. When irrigation time is determined, on the last stage we must obtain the value of irrigation rate that is the volume of water needed to be supplied. We use two approaches to determine it. Within the first one, while modeling moisture transport according to the model (1), we fix the irrigation flux Qi from the specific irrigation machine and calculate the estimated time needed to increase moisture content in soil’s root layer up to the given upper limit. Multiplying this time on Qi we obtain an irrigation rate. Such approach take into account water percolation below root layer but is quite sensible to the accuracy of irrigation flux measurements and errors of numerical method used to solve initial-boundary value problem for the Eq. (1). Thus, we also use the second approach based on the rather simple formula of A.N. Kostiakov. Summarizing, the input data for the considered modeling base are discretization steps for the differential model; water head measurements; surface air temperature and humidity measurements along with their forecasts; precipitation measurements and forecast; actual irrigation times and rates; soil’ water retention curves and dependencies of hydraulic conductivity on water head; optionally, plant’s root system development parameters; range in which moisture content should be maintained for the specific crop in the specific growing conditions and development stage. With such inputs, a decision support algorithm that generates recommendations on the time and rate of the next irrigation for a single field can be stated as follows: • Check if the model (1) and its classical analogue were previously been calibrated. Perform re-calibration using water head measurements for a given period t 1 preceding the current moment if within a given period t 2 the best model describes the measured data with an average relative error more than a given threshold ε; • Perform models’ calibration if they were not previously been calibrated and the needed amount of input data are collected; • Perform predictive modeling of water head dynamics for a given period t 3 using the fractional differential model (1), its classical analogue, and the balance model;

Decision Support System in Sprinkler Irrigation

19

• Determine the next irrigation time using the model that best describes the measurements within the period t 1 preceding the current moment of time. Determine the irrigation rates using the two above-described approaches. The choice of the recommended rate should be performed by expertise. Generate a recommendation.

4 Software Implementation The proposed algorithm was implemented in the irrigation management system “Irrigation Online” and is currently accessible through http://185.168.130.174:90/. The system’s database is organized within the PostgreSQL database management system with PostGIS extension for storage and processing of georeferenced data. The database consists of tables in which the following is stored: • information about monitoring network: serviced organizations (farms), agricultural fields, micro weather stations and their location; • information on crops and their biological characteristics, including information needed to manage irrigation depending on the stage of crop development; • data on hydrophysical properties of soil profiles such as water retention curves in the form of van Genuchten model’s coefficients and filtration coefficients; • monitoring data, weather forecasts, simulation results, and recommendations for irrigation generated on the base of simulation results. The iMetos® or Davis® weather stations are used as a hardware in the system. The system performs the monitoring of moisture content and the availability of soil moisture for plants by Watermark® Irrometer™ sensors [32]. Other sensors necessary for system’s operation are the sensors of temperature and humidity of surface air; precipitation sensor. Weather station sensors’ readings are stored in the database on hourly basis using the modules that interact with stations’ APIs. The module that interacts with Davis® stations via the HTTPS API and is implemented in Python language is called once an hour to retrieve current sensors’ readings and append them into the database. The modules for iMetos® stations use interface software to the appropriate APIs implemented in PHP and provided by the station’s developer. They consist of PHPimplemented programs that retrieve retrospective data from the station and once-a-day called Python-implemented program that writes the data into the database. For both modules, the station’s access parameters, the station ID within the database, and the matching between sensor names and the fields of the corresponding database table are specified in separate text configuration files. In addition to the actual data coming from weather stations located in fields, the system needs some predicted data that are automatically obtained from weather forecasting services [33]. Given the high computational complexity of moisture transfer modeling, the solution of initial-boundary value problems for the Eq. (1) was implemented in the module mt written in C++ language using OpenMP multi-threading. This module performs both model’s calibration and prediction requiring input data described in the text file that contains

20

V. Bohaienko et al.

• readings of water head sensors in the time periods used to calibrate the model and verify its accuracy; • water head sensors’ readings at the initial time used for calibration or prediction; • precipitation amount and actual rates of conducted irrigations. In the case of model’s calibration, actual values should be given. Forecast of precipitation can be used for prediction; • evapotranspiration values calculated on the base of actual or predicted surface air variables. The values of model’s parameters resulting from its calibration are stored in a text file and are used to further predict the dynamic of water head at the depths of sensors’ placement. The results of prediction are also written into a text file. Performing simulations, the discretization steps are equal to 1 cm with respect to the depth variable and to 5 min with respect to the time variable. The Python-implemented module mt_run.py was created to extract for a specific weather station the necessary data from the database; basically verify their adequacy; calculate evapotranspiration; generate an input data file; and run the mt module. When specified by the module’s flags, mt_run.py checks the consistency of the previously computed predictions with the actual data obtained after its generation and executes the mt module to re-calibrate the model if the accuracy is less than a given acceptable level. It further performs water head dynamic prediction using either new or previously obtained optimal values of models’ parameters. After performing the predictive simulation, the resulting data from the generated text file is parsed and appended into the database. The separate once-a-day run run_forecast shell script executes the mt_run.py module using shell configuration file for all the weather stations managed by the system. This script also runs the re-calibration procedure once a week. Once a day, the prediction of water head dynamic for the next 5-day period is performed. Recommendations (time and rate of the next irrigation) for the farm fields in the serviced areas are generated on its base. For recommendations generation, we choose the model that most accurately described water head dynamic for the last 24 h. For each of the serviced fields, the system generates recommendations of three types with the corresponding irrigation rates: “irrigation is urgent”, “irrigation is not needed in the next 5 days”, and “irrigation must be performed within some specified time in the next 5 days”. The “irrigation is urgent” recommendation is provided in cases when user delayed irrigation for various reasons or when irrigation needs to be performed today according to the actual sensors’ readings. Recommendation generation is performed by the Python-implemented module that is run once a day after launching iMetos® weather data acquisition module and the run_forecast script that performs soil moisture content prediction. The developed DSS gives users the opportunity to get information about the current and future state of soil moisture at several levels of detail as • the list of charts on water head dynamic and current recommendations for each field of a farm; • charts of actual values of the most important indicators (water heads, temperature and humidity, precipitation) for a specific field and weather station;

Decision Support System in Sprinkler Irrigation

21

• charts of actual and predicted values of all indicators stored within the system’s database, including calculated values of evapotranspiration, meteorological forecasts, and predictions of water head dynamic by different models. The system contains tools for its administration, in particular, for the entry of hydrophysical parameters of soils and irrigation parameters in accordance with the stages of crop’s development. All tools of the system are accessible from within the interactive map. The GUIs are implemented in PHP using AJAX POST requests to modify the system’s database. The interactive map’s implementation uses the OpenLayers library and OpenStreetMaps maps. At the current stage of development the only software’s interface’s locale is Ukrainian.

5 Experimental Assessment and Practical Application The testing of the mathematical model and the algorithm of its parameters identification, which form the core of the presented decision support system, was performed in 2018 using the monitoring data collected in the fields of the State Enterprise Farm “Brilivske” (Pritvitne village, Oleshkivsky district, Kherson region, Ukraine) during soybeans growing under irrigation by a center pivot sprinkling machine. Point measurements of water heads were performed by the Watermark sensors placed at 6 different depths. The sensors readings were collected once per hour. For 3 month modelling period, the classical one-dimensional Richards equation’s usage results in the highest relative error equal to 9.5% [25] while the usage of the fractional differential model allowed improvement of the simulation accuracy by 7%. The simulated seasonal irrigation rate in this case differed from the actual one not more than by 12%. Further experimental studies [30] based on the dataset collected in 2019 showed the possibility of additional accuracy improvement introducing more general operators into the Eq. (1). In 2020, the presented algorithm and software was used to schedule irrigation in 6 fields of the “Adelaida” farm in Kherson region of Ukraine. Potato was grown in 5 fields and wheat in one field. Compared to the previous year, the usage of the software allowed farmers to use irrigation water more efficiently, to apply up to 30% less water obtaining equal or even higher yields The average water savings reached 1,000 m3 /ha.

6 Conclusions The paper presents a novel algorithm of a decision support system in irrigated agriculture that combines the usage of the generalized fractional Richards equation stated in terms of water head to predict soil moisture content dynamic and metaheuristic particle swarm optimization technique to perform the identification of the model’s parameters. Such approach allowed obtaining up to 7% accuracy increase while modeling water head dynamics compared to the usage of an integer-order model.

22

V. Bohaienko et al.

Moreover, compared to the known decision support systems, we propose to perform parameters re-calibration on the base of actual sensors readings as errors accumulate. This allows transparently taking into account both the changes in soil structure and crop coefficient during a growth season that in turn results in the improvement of recommendations’ quality. The modular structure of the system allows it to be expanded from the standpoint of the increase in the diversity of its monitoring data sources and mathematical modeling tools. In particular, further software implementation of two-dimensional models of moisture transport, e.g. those used in [34], will enable to use the system for supporting the decision-making in drip irrigation without the significant modification of other components. Regarding high computational complexity of solving fractional differential equations further works can also be performed in the field of parallel algorithms’ [35, 36] application to increase system’s reliability under high load conditions.

References 1. Akwu, S., Bature, U.I., Jahun, K.I., Baba, M.A., Nasir, A.Y.: Automatic plant irrigation control system using Arduino and GSM module. Int. J. Eng. Manuf. 10(3), 12–26 (2020). https://doi. org/10.5815/ijem.2020.03.02 2. Okine, A., Appiah, M., Ahmad, I., Asante-Badu, B., Uzoejinwa, B.: Design of a green automated wireless system for optimal irrigation. Int. J. Comput. Netw. Inf. Secur. 12(3), 22–32 (2020). https://doi.org/10.5815/ijcnis.2020.03.03 3. Rinaldi, M., He, Z.: Decision support systems to manage irrigation in agriculture. Adv. Agron. 123, 229–279 (2014). https://doi.org/10.1016/B978-0-12-420225-2.00006-6 4. Pachepsky, Y., Timlin, D.: Water transport in soils as in fractal media. J. Hydroogy 204(1–4), 98–107 (1998). https://doi.org/10.1016/S0022-1694(97)00110-8 5. Pachepsky, Y., Timlin, D., Rawls, W.: Generalized Richards’ equation to simulate water transport in unsaturated soils. J. Hydrol. 272, 3–13 (2003). https://doi.org/10.1016/S00221694(02)00251-2 6. Tu, T., Ercan, A., Levent Kavvas, M.: Time–space fractional governing equations of transient groundwater flow in confined aquifers: numerical investigation. Hydrol. Process. 32, 1406– 1419 (2018). https://doi.org/10.1002/hyp.11500 7. Romashchenko, M.I., Matiash, T.V., Bohaienko, V.O., Kovalchuk, V.P., et al.: Development experience and ways of improvement of irrigation management systems (in Ukrainian). Land Reclam. Water Manag. 2, 17–30 (2019). https://doi.org/10.31073/mivg201902-207 8. Steduto, P., Hsiao, T.C., Raes, D., Fereres, E.: AquaCrop—the FAO crop model to simulate yield response to water: I Concepts and underlying principles. Agron. J. 101, 426–437 (2009). https://doi.org/10.2134/agronj2008.0139s 9. Doorenbos, J., Kassam, A.H.: Yield Response to Water. FAO Irrigation and Drainage Papers No. 33. FAO, Rome (1979) 10. Steduto, P., Raes, D., Hsiao, T.C., Fereres, E.: AquaCrop: concepts, rationale and operation. In: Steduto, P., Hsiao, T.C., Fereres, E., Raes, D. (eds.) Crop Yield Response to Water.FAO irrigation and drainage paper no. 66, pp. 17–49. FAO, Rome (2012) 11. Allen, R.G., Pereira, L.S., Smith, M., Raes, D., Wright, J.L.: FAO-56 dual crop coefficient method for estimating evaporation from soil and application extensions. J. Irrig. Drainage Eng. 131(1), 2–13 (2005). https://doi.org/10.1061/(ASCE)0733-9437(2005)131:1(2) 12. Shtoiko, D.A., Pysarenko, V.A., Bychko, O.S.: Estimated methods for determining total evaporation and irrigation time of crops. Zroshuvalne zemlerobstvo, pp. 3–8 (1977). (in Ukrainian)

Decision Support System in Sprinkler Irrigation

23

13. Ivanov, N.N.: On the Determination of Evaporation Values, pp. 189–196. Yzv. HHO, Moskow (1954). (in Russian) 14. Williams, J.R., Izaurralde, R.C.: The APEX model. BRC Report 2005-02. Blackland Research and Extension Center, Blackland (2005) 15. Borah, D.K., et al.: Sediment and nutrient modeling for TMDL development and implementation. Trans. ASABE 49(4), 967–986 (2006) 16. Panagopoulos, Y., Makropoulos, C., Mimikou, M.: Decision support for diffuse pollution management. Environ. Model Softw. 30, 57–70 (2012). https://doi.org/10.1016/j.envsoft.2011. 11.006 17. Styczen, M., Poulsen, R.N., Falk, A.K., Jørgensen, G.H.: Management model for decision support when applying low quality water in irrigation. Agric. Water Manag. 98, 472–781 (2010). https://doi.org/10.1016/j.agwat.2010.10.017 18. Smith, M.: CROPWAT, a computer program for irrigation planning and management. FAO Irrigation and Drainage Paper No. 46. (1992) 19. Car, N.J., Christen, E.W., Hornbuckle, J.W., Moore, G.A.: Using a mobile phone short messaging service (SMS) for irrigation scheduling in Australia—farmers’ participation and utility evaluation. Comput. Electron. Agric. 84, 132–143 (2012). https://doi.org/10.1016/j.compag. 2012.03.003 20. Abrahamsen, P., Hansen, S.: Daisy: an open soil-crop-atmosphere model. Environ. Model Softw. 15, 313–330 (2000). https://doi.org/10.1016/S1364-8152(00)00003-7 21. Zhang, Y., Feng, L.: CropIrri: a decision support system for crop irrigation management. In: Li, D., Zhao, C. (eds.) CCTA 2009. IAICT, vol. 317, pp. 90–97. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-12220-0_14 22. Keating, B.A., et al.: An overview of APSIM, a model designed for farming system simulation. Eur. J. Agron. 18(3), 267–288 (2003). https://doi.org/10.1016/S1161-0301(02)00108-9 23. Stockle, C.O., Donatelli, M., Nelson, R.: CropSyst, a cropping systems simulation model. Eur. J. Agron. 18, 289–307 (2003). https://doi.org/10.1016/S1161-0301(02)00109-0 24. van Genuchten, M.T.: A closed-form equation for predicting the hydraulic conductivity of unsaturated soils 1. Soil Sci. Soc. Am. J. 44(5), 892–898 (1980) 25. Romashchenko, M.I., Bohaienko, V.O., Matiash, T.V., Kovalchuk, V.P., Danylenko, Iu.Iu.: Influence of evapotranspiration assessment on the accuracy of moisture transport modeling under the conditions of sprinkling irrigation in the south of Ukraine. Arch. Agron. Soil Sci. 66(10), 1424–1435 (2020). https://doi.org/10.1080/03650340.2019.1674445 26. Rosetta Version 1.0 (Free downloaded program). U.S.Salinity Laboratory ARSUSDA. http:// www.ussl.ars.usda.gov. Accessed 10 Sept 2020 27. Bogaenko, V.A., Bulavatsky, V.M., Kryvonos, Iu.G.: On mathematical modeling of fractionaldifferential dynamics of flushing process for saline soils with parallel algorithms usage. J. Autom. Inf. Sci. 48(10), 1–12 (2016). https://doi.org/10.1615/JAutomatInfScien.v48.i10.10 28. Bulavatsky, V.M.: Mathematical modeling of dynamics of the process of filtration convection diffusion under the condition of time nonlocality. J. Autom. Inf. Sci. 44(2), 13–22 (2012). https://doi.org/10.1615/JautomatInfScien.v44.i4.20 29. Zhang, Y.A.: Comprehensive survey on particle swarm optimization algorithm and its applications. Mathematical Problems in Engineering, Article no 931256 (2015). https://doi.org/ 10.1155/2015/931256 30. Bohaienko, V., Gladky, A., Romashchenko, M., Matiash, T.: Identification of fractional water transport model with ψ-Caputo derivatives using particle swarm optimization algorithm. Appl. Math. Comput. 390, Article no 125665 (2021). https://doi.org/10.1016/j.amc.2020. 125665 31. Rao, N.H.: Field test of a simple soil-water balance model for irrigated areas. J. Hydrol. 91, 179–186 (1987). https://doi.org/10.1016/0022-1694(87)90135-1

24

V. Bohaienko et al.

32. IRROMETER Company Inc. https://www.irrometer.com/sensors.html. Accessed 10 Sept 2020 33. Kovalchuk, V., Demchuk, O., Demchuk, D., Voitovich, O.: Data mining for a model of irrigation control using weather web-services. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2018. AISC, vol. 754, pp. 133–143. Springer, Cham (2019). https://doi.org/10. 1007/978-3-319-91008-6_14 34. Romashchenko, M.I., Shatkovsky, A.P., Onotsky, V.V.: Mathematical model of flat-vertical profile moisture transfer under trickle irrigation in conditions of incomplete saturation. Agric. Sci. Pract. 3(3), 35–40 (2016). https://doi.org/10.15407/agrisp3.03.035 35. Bohaienko, V.O.: Parallel finite-difference algorithms for three-dimensional space-fractional diffusion equation with ψ-Caputo derivatives. Comput. Appl. Math. 39(3), 1–20 (2020). https://doi.org/10.1007/s40314-020-01191-x 36. Bohaienko, V.O.: Parallel algorithms for modelling two-dimensional non-equilibrium salt transfer processes on the base of fractional derivative model. Fract. Calculus Appl. Anal. 21(3), 654–671 (2018). https://doi.org/10.1515/fca-2018-0035

Automated Pipeline for Training Dataset Creation from Unlabeled Audios for Automatic Speech Recognition O. Romanovskyi1

, I. Iosifov1

, O. Iosifova1 , V. Sokolov2(B) and I. Sukaylo2

, F. Kipchuk2

,

1 Ender Turing OÜ, Tallinn, Estonia

{or,ei,oi}@enderturing.com 2 Borys Grinchenko Kyiv University, Kyiv, Ukraine

{v.sokolov,f.kipchuk.asp,i.ukailo.asp}@kubg.edu.ua

Abstract. In the paper, we present a software pipeline for speech recognition to automate the creation of training datasets, based on desired unlabeled audios, for low resource languages and domain-specific area. Considering the commoditizing of speech recognition, more teams build domain-specific models as well as models for local languages. At the same time, lack of training datasets for low to middle resource languages significantly decreases possibilities to exploit last achievements and frameworks in the Speech Recognition area and limits the wide range of software engineers to work on speech recognition problems. This problem is even more critical for domain-specific datasets. The pipeline was tested for building Ukrainian language recognition and confirmed that the created design is adaptable to different data source formats and expandable to integrate with existing frameworks. Keywords: Automatic Speech Recognition · ASR · Dataset creation pipeline · Natural language processing · NLP · Asynchronous graphs

1 Introduction Speech assistants have become more widely used by people [1–3]. It leads to more corporations, startups, and research teams ready to build Automatic Speech Recognition (ASR) models. This readiness faced major limitation: lack of available datasets to train ASR models, especially critical for low resource languages and specific domain of desired ASR model (e.g., medical, finance, insurance, etc.). This limitation also breaks attracting software engineers to the speech recognition area. In order to take away such limitation, teams have to accomplish multiple steps: get a set of relevant audios; get or create transcription for all or part of it (most comprehensive and costly operation); clean up the data; convert it to format, supported by a tool for model training; train a model [4]. Big companies mostly publish research papers and open-source projects that are focused on training models that can achieve a lower error rate by using one of the public © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 25–36, 2021. https://doi.org/10.1007/978-3-030-80472-5_3

26

O. Romanovskyi et al.

datasets for training and validation [5, 6]. On the other hand, building an efficient pipeline for preparing datasets is a critical part of getting a high-quality ASR model. So, there is a need for a tool that either covers pre-training steps or all steps. In this paper, we designed and implemented an ASR software pipeline framework for training dataset formation approach from raw audios (for desired language or domainspecific unlabeled audios). It supports lazy computation, transformation results caching, automatic full and partial cache invalidation. We make the following contributions in this: (1) designed and implemented the ASR pipeline framework, as well as some of a subset of commonly used data transformations; (2) defined data structures to support interoperability between transformations, created by different developers; (3) conducted experiments prove that the proposed caching approach improves reprocessing speed; (4) was used the implemented framework to build a dataset for the Ukrainian language that is bigger than known private datasets “Ukrainian Corpus for Broadcast Speech” of 366 h [7]. The main objectives for such a pipeline were to decrease the time consumption of the researcher or developer to prepare a training dataset of the desired size and decrease the learning curve for newbies to start working with available ASR frameworks and tools. Section 2 provides an analysis of the sources. Sect. 3 contains the research methodology and indicates the direction of research. Comparison with related tools and frameworks are described in Sect. 4. Design process and implementation details are shown in Sect. 5. Results of conducted experiments are given in Sect. 6. The paper ends with acknowledgments in Sect. 7 and conception of the future work in Sect. 7.

2 Review of the Literature Because the quality of the trained model depends on the amount of audio data, a quantity of audio data is very important. But the amount of data is very costly. The paper [8] solving the problem of insufficient data by increasing the dataset to 53,000 h, which saves a lot of resources for work. But this approach requires pre-training models by masking part of data with time steps of output data, preparing the pipeline for unlabeled data, and some resources. The work [9] is dedicated to comparing training approaches to supervised and unsupervised data. Exploring the different training language models show the real performance of exact speech recognition methods. The proposed strategy is to train the next hour of unlabeled data by one hour of manually trained data and then repeating this operation by doubling. Unsupervised data requires a huge amount of unlabeled audio and the corresponding resources for preparation and training models which are absent in Ukrainian and were a reason to choose supervised data. The practical usage of the paper [10] is a good example of using one of the biggest resources—Youtube.com to implement ASR for training deep neural networks and creating video subtitles from audio data. Making correct text alignment to audio is calculated by the “confidence” factor, which partially relies on manual training. Using semi-supervised auto-sync data speech recognition shows better performance and is important to make correct speech recognition and alignment for unsupervised media data.

Automated Pipeline for Training Dataset Creation

27

Our previous studies have looked at techniques comparison for natural language processing [11] and sentence segmentation from unformatted text using language modeling and sequence labeling approaches [12].

3 Research Methodology When starting a new model, one of the major limiting factors is time. It can be either direct limitation (available time for research), or indirect limitation (the period covered by investments to develop a product). We identified two factors that can be optimized. The first factor is the time to start. Existing tools have a steep learning curve (e.g. Kaldi) and assume the existing dataset in a predefined format (e.g. wav2letter). From our experience and discussion with the community, it’s not uncommon to spend days and, in some cases, weeks before the start of the first model training. If the first factor has to be done once, the second factor is a recurring one. Creating a custom audio dataset is an ongoing process with adding new data, removing bad-quality audios, etc. In such a scenario, reprocessing all audios after each change slows down the experimenting rate. If we can reprocess only the modified part—it would allow doing more experiments and will lower computation time and power requirements.

4 Comparison with Related Tools and Frameworks 4.1 Defining Criteria to Compare We defined the minimal set of requirements for a pipeline framework to address the slowing factors from Sect. 3: (1) support building of preprocessing pipeline; (2) support for data from different sources in different formats; (3) support persisting and invalidation of outputs of preprocessing steps; (4) first-class integration with text and audio transformations; (5) low learning curve. To make it formal we define a set of factors that would lower the learning time: (1) single programming language (some tools use a mix of different languages, like Bash, Perl, Python, C++ but it increases environment setup time); (2) support for ASR-specific routines (for audio and text) out of the box or having a documented way of integrating them; (3) modern Integrated Development Environment (IDE) support. The above list can be used as criteria to evaluate existing tools as well as a set of requirements for the designed pipeline framework. 4.2 Existing Tools In Table 1 we presented a comparative analysis of the most commonly used existing SPT (Speech Processing Toolkit) and tools to build an ASR model, including NLP (Natural Language Processing) tools not directly related to ASR, but related to data preprocessing, like TTK (Text Toolkit) Huggingface for normalization and DLF (Deep Learning Framework) Tensorflow/Pytorch for grapheme to phoneme model training (see Table 1).

28

O. Romanovskyi et al. Table 1. Comparative analysis of existing software.

Product

Type

Training type

Data amount

ASR-specific routines

Cache

Learning curve

Kaldi

ASR toolkit Hybrid SPT

Supervised

Low-Medium

Wide variety of audio and text data preprocessing scripts; training routines

Manual reusing of results, manual invalidation

Hard

Julius

ASR toolkit Hybrid SPT

Supervised

Low-Medium

Relies on external tools for data preprocessing

Not supported

Hard

DeepSpeech

ASR toolkit End-to-End SPT

Supervised

Medium-High

LM building functions and feature extraction routines for audio

Only supported for MFCC

Easy

EspNet

ASR toolkit End-to-End SPT

Supervised

Medium-High

Wide variety of audio and text data preprocessing scripts; training routines

Manual reusing of results, manual invalidation

Easy

Wav2Vec

ASR toolkit End-to-End SPT

Unsupervised

Low-High

Wide variety of audio and text data preprocessing routines

Provided utils for in-memory caching

Normal

Hugging Face

Text toolkit PyTorch, TensorFlow

Supervised/ unsupervised

Text tokenizers and normalization

Supported, no partial invalidation

Easy

—

Hard

—

Wide variety of audio and text data preprocessing routines in “torch-” text & audio modules

Not supported

—

—

A small set of text preprocessing routines, a wide variety of signal processing routines for audio

Supported via snapshot API, no partial invalidation

Hard

—

Pytorch

TensorFlow

DLF

DLF

As can be seen from comparative analysis, no one toolkit can handle all steps of ASR data-preprocessing (including collection, splitting, labeling, preparing for ASR expected format), and model training.

5 Design Process and Implementation Details As was defined in Sect. 4.1, using a single programming language is a factor to lower the learning curve. Moreover, we assume that data scientists and machine learning engineers are likely to be part of the ASR team; it would be beneficial to use a programming language they are familiar with. As a result, we decided to use Python for the pipeline framework. We decided to focus on small and medium-sized datasets (up to singledigit thousands of hours). Having a data size that can be handled on a single machine allows us to simplify the pipeline usage. We consider it an acceptable trade-off since new researchers and teams are unlikely to have more than 10,000 h of audio.

Automated Pipeline for Training Dataset Creation

29

5.1 Main Components As shown in Fig. 1, we have defined four main types of components: • DataProxy. An object to pass data between pipeline steps. • Transformations. Are processing components that take one or more DataProxy objects and return new DataProxy. The concept is similar to Tensorflow Transformations [13]. • Data Input. Used to read data from external sources. Unlike Transformations, they don’t take DataProxy as input but do produce DataProxy output. A most common use case would be reading source audio files and transcripts. • Data Output. Used to export results of processing in a certain format. Unlike Transformations, they take DataProxy as input but do not produce DataProxy output. The most common usage scenario is saving a dataset in a format of a certain model-training toolkit (e.g. creating an input dataset for Kaldi or Wav2Vec). This component allows conducting experiments with different model architectures and training tools while reusing the same preprocessing pipeline.

Fig. 1. Main pipeline components.

A combination of the defined components allows us to define pipelines. As a next step, we define data structures and API for them to achieve interoperability between components. 5.2 Data Structures Selection Before defining data structures for components, we need to decide how to connect separate components into a single pipeline. As can be seen in other computational frameworks [13–15] one of the common methods of creating complex computation from multiple components is using directed acyclic graphs. It fits ASR pipelines as well, moreover, using concepts that people can be familiar with also lowers learning time. Therefore, we also represent computations structure in a form of a directed acyclic graph as shown in Fig. 2. To make different components compatible with each other, we have to define the data format to be passed between them. Data structure would include two major components: main data (e.g., transcripts and references to audio files); metadata (e.g., caching information). Metadata itself can consist of two parts. One part is transformation-level data (e.g., when transformation started/finished). The second part of the metadata is row-level. It would store meta information for each audio file or utterance. Row-level data is needed to support partial invalidation. Metadata is covered in detail in Sect. 5.4. As the base structure, Suggest uses Pandas DataFrame. It has two features that are useful for creating an ASR pipeline: (1) built-in support for table-level operations, which

30

O. Romanovskyi et al.

Fig. 2. Proposed data structure components by example.

allows writing more concise and clean code; (2) there are frameworks for distributing DataFrame computations like Modin [16] in case we decide to handle bigger datasets in the future. A recommended set of columns for DataFrame: • Audio file URI. As in the first version, we only support the local file path, but in the future, we add support for other URI schemas. • Utterance start time. Defines where an utterance starts. If the value is not specified— we assume that the file has a single utterance. • Utterance end time. Defines where an utterance ends. If the value is not specified— we assume that the file has a single utterance. • Text. A transcript of the utterance. Can be not set if further transformations would add it. • Tag. Tag is an optional data label. For example, it can be used to set a data source to be able to track data through the pipeline. • Subset. Defines if a specific record belongs to the train set, test set, or validation set. Another option would be to have two different objects. One for the train set and one for the test set. But since all preprocessing is the same for both, having them in a single DataFrame simplifies the code. Also having the column from the start gives the flexibility of how to split data. E. g. we can have a small subset of high-quality data to be used as a validation set. Then the column would be set from “Data Input” components. Or we can have a random split as one of the transformations. Defines if a specific record belongs to the train set, test set, or validation set. Another option would be to have two different objects. One for the train set and one for the test set. But since all preprocessing is the same for both, having them in a single DataFrame simplifies the code. Also having the column from the start gives the flexibility of how to split data. All transformations pass-through columns they don’t modify. It allows creating single-responsibility transformations that don’t break other parts of the data. 5.3 Programming Interfaces Unifying implementations of transformations gives the following advantages: (1) shortens learning time since a user only learns once how to call each transformation and how to combine multiple transformations; (2) allows having a standard implementation of caching that can be reused by different transformations; (3) it makes components

Automated Pipeline for Training Dataset Creation

31

more future proof since we can add a global task scheduler for distributed computations without rewriting transformations from scratch. We considered three ways of structuring transformations: inheritance, strategy, and decorator design patterns. With inheritance, we can implement a Template Method design pattern. The simplified structure for the base class would look like this: class TransformationBase(ABC): @abstractmethod def _process(self, inp, **kwargs): # would be overridden in transformations pass def process(self, inp, **kwargs): # validation and caching can go here result = self._process(inp, **kwargs) # updating cache can go here return result

Then transformations implement the _process method. But this approach wouldn’t allow using IDE autocomplete and type checks for “process” method parameters. The strategy design pattern also doesn’t give the ability to utilize IDE auto-complete. It’s mostly used to modify an object’s behavior in some aspects, but in the case of transformations, we change the main part of behavior, so it can be misleading for users. Decorator design pattern allows us to augment the behavior of transformations by adding common functionality like caching, logging, time measuring, graph checks, etc. Implementation of a transformation with a decorator can look like: @transformation class MergeStreams: def process(self, data_frames: list[pd.DataFrame]): res = pd.concat(data_frames).reset_index() log.info("Merged %d sources, got %d records.", len(data_frames), len(res)) return res

Having a decorator and a computation graph gives an additional advantage—we can check whether results of a specific transformation are used and only compute what’s needed. Below is an example of a pipeline that uses transformations, defined, and augmented with decorators: data = MergeStreams().process([ get_audio_books_stream(), get_calls_stream(), ]) data = NormalizeNumbers().process(data) data = TrainTestSplit().process(data) KaldiOutput(dst_dir=data_dst, sample_rate=16000).process(data).compute()

32

O. Romanovskyi et al.

The “compute” method at the end of the code explicitly starts computations since it’s the last item of the pipeline. 5.4 Automatic Caching and Invalidation Automatic caching and invalidation allow speed-up experiments, reprocessing source data with some additional steps, etc. Caching includes the following decisions: what transformations can and should be cached; how to check if data is not changed; how to track which part of data was changed to be able to do partial invalidation; how to store data. Below we cover those decisions. Considering the target domain (working with custom datasets)—the common scenario would be adding or removing part of the data. In such cases, the caching routine has to detect a modified part of the data and only do calculations on a changed part. To be able to do it at all steps of a pipeline we need a way of mapping DataInput row to a row or a set of rows in the output of further rows. Let’s take a pipeline, consisting of the following steps: 1. 2. 3. 4. 5.

DataInput from *.csv and *.mp3. Normalize numbers. Split long segments into shorter segments. Shuffle and do train/test split. Save in ASR training framework format.

In a given pipeline, removing a single source file will cause the removal of a single row in Step 2, removing one or more rows in Step 3, removing the same number of rows but in different positions in Step 4, different output in Step 5. To be able to track rows over multiple transformations we need a way to identify each row. Since we don’t have unique ids, we suggest using a hash of all DataFrame columns as row id for a transformation. As a result, to support cache invalidation, including partial invalidation, a list of DataFrame columns, defined in Sect. 5.2 would be extended by two columns: hash, parent_hash. On the output level, we store three hash values: __init__ parameters hash; process parameters hash; audio files hash.

6 Experiments 6.1 Pipeline Implementation As we aimed to create a pipeline and use it with real-life tasks, we have chosen the Ukrainian language as a candidate language to apply the created pipeline. Because there are too few publicly available datasets (267 h) to achieve good results of the STT model without collecting and building a training dataset. Table 2 presented publicly available ASR datasets for Ukrainian (low resource language), as collecting such publicly available datasets is the starting point for each ASR model training. Such an amount of data is not enough for a qualitative ASR model. By applying the current pipeline, we managed

Automated Pipeline for Training Dataset Creation

33

to create 2,500 h of training ASR dataset for 84 h which leads us to the state-of-theart Word Error Rate (WER) of 5.24 for the Ukrainian language-based using Mozilla Common Voice (22 h) as validation dataset. Table 2. Datasets and Sources for Ukrainian Speech Corpus. Dataset name

Class

Duration, hours Speakers Quality

Ukrainian Corpus for Broadcast Speech [17, 18]

ASR dataset 366

Multi-speaker Corpus “UkReco” [19]

ASR dataset

M-AILABS Ukrainian Corpus [20]

Books

330

Publicly available No

— 0.

as k > 1.

Variance of such random variable is Dξ =



xmin k−1

2

k k−2 ,

as k > 2.

To simulate ξ we can use an inverse function method. Then ξ = uniformly distributed on (0; 1).

xmin

1

(1−α) k

, where α is

Estimation of Hurst Index and Traffic Simulation

43

Figure 2 shows the realizations of random variables following Pareto distribution with k = 3, xmin = 2. The estimates of the expected value and variance of random variable with Pareto distribution are E ξˆ = 3.05, Dξˆ = 3.245 respectively in the case when k = 3, xmin = 2.

Fig. 2. Realizations of random variable following Pareto distribution.

Hypothesis testing of Pareto distribution is accepted at the level ε = 0, 95. Weibull Distribution. Weibull distribution has in such form     x k k > 0, λ > 0. F(x) = 1 − exp − λ   The mean of random variables following Weibull distribution is Eξ = λ 1 + k1 .      Variance equals Dξ = λ2  1 + k2 −  2 1 + k1 . To simulate Weibull distribution the inverse function method can be used. Then 1 ξ = λ[− ln(1 − α)] k , where α is uniformly distributed on (0; 1). Figure 3 shows the realizations of random variable following Weibull distribution with k = 3, λ = 2.

Fig. 3. Realizations of random variable following Weibull distribution.

The estimates of expected value and variance of random variable with Weibull distribution with k = 3, λ = 2 are E ξˆ = 1.787, Dξˆ = 0.421 respectively.

44

A. Pashko et al.

Caushy Distribution. The Caushy distribution has cdf   1 x − x0 1 F(x) = arctg + , γ > 0, x0 ∈ (−∞; +∞). π γ 2 The mean of random variable with Caushy distribution doesn’t exist. The variance is Dξ = +∞. In the case when γ = 1 and x0 = 0 the distribution is called standard Caushy distribution. To simulate random variable distribution we can use an inverse   with Caushy function method. Then ξ = x0 + γ tg π x − 21 , where α is uniformly distributed on (0; 1). Figure 4 shows the realizations of random variables following Caushy distribution with γ = 1, x0 = 0.

Fig. 4. Realizations of random variables following Caushy distribution.

The estimates of the expected value and variance of random variable with log-normal distribution are E ξˆ = 1.104, Dξˆ = 5.06 ∗ 104 respectively in the case when γ = 1, x0 = 0. Caushy distribution describes amplitude-frequency characteristics of linear oscillation systems in the neighborhood of resonance frequencies. If η1 , η2 are independent random variables with N(0, 1) distribution then a random variable ηη21 has standard Caushy distribution. Nakagami Distribution. The Nakagami distribution (or Nakagami-m distribution) has probability density function  m  1 2  m m 2m−1 x exp − x2 m ≥ , w > 0. f (x) = (m) w w 2 The expected values of random variable with Nakagami distribution is Eξ = ⎛    2 ⎞ 1  m+ 2 1 w 2 1 ⎠. ⎝ (m) m . The variance is Dξ = w 1 − m (m)

   m+ 12 

To simulate random variable with Nakagami distribution the inverse function method isn’t applied explicitly. In case of integer m the following representation can be used for simulation     m  1    ξ = ηi2 + ζi2 , 2m i=1

Estimation of Hurst Index and Traffic Simulation

45

where {ηi , ζi }m i=1 are the sequence of independent random variables with N(0; w) distribution. Figure 5 shows the realizations of random variables following Nakagami distribution.

Fig. 5. Realizations of random variable following Nakagami distribution.

The parameters of Nakagami distribution are calculated in a such way m =   w = E ξ2 .

  E2 ξ 2 , D (ξ 2 )

4 Conclusions In the paper, we studied the methods of statistical simulation of the traffic and estimation of Hurst index with measurement errors. The main distributions were also considered, which are used to create “Internet of Things” traffic. Investigating the multiservice traffic it should be taken into account the possible change in its nature as the number of end devices through which the concept of “Internet of Things” is implemented. In the coming years, the volume of IP packets related to Internet of Things traffic will be increased. A special interest in the study of multiservice traffic is paid to the distributions of interarrival times IP packets, which are determined at the finite time interval. The study of such distributions provides a new class of problems in the theory of teletraffic with practical significance. In future, we will apply the constructed estimator of Hurst index for the real data and simulate traffic using FBM models.

References 1. Messier, G.G., Finvers, I.G.: Traffic models for medical wireless sensor networks. IEEE Commun. Lett. 11(1), 13–15 (2007). https://doi.org/10.1109/LCOMM.2007.061291. https:// ieeexplore.ieee.org/document/4114210 2. Wang, Q., Zhang, T.: Source traffic modeling in wireless sensor networks for target tracking. In: Proceedings of the 5th ACM International Symposium on Performance Evaluation of Wireless Ad-Hoc, Sensor, and Ubiquitous Networks (PE-WASUN 2008), pp. 96–100 (2008) 3. Vybornova, A., Koucheryavy, A.: Traffic analysis in target tracking ubiquitous sensor networks. In: Balandin, S., Andreev, S., Koucheryavy, Y. (eds.) Internet of Things, Smart Spaces, and Next Generation Networks and Systems, pp. 389–398. Springer, Cham (2014). https:// doi.org/10.1007/978-3-319-10353-2_34

46

A. Pashko et al.

4. Levakov, A.K., Sokolov, A.N., Sokolov, N.A.: Models of incoming traffic in packet networks. T-Comm 9(5), 91–94 (2015) 5. Hosseini, S.A., Akbarzadeh-T, M.-R., Naghibi-Sistani, M.-B.: Qualitative and quantitative evaluation of EEG signals in epileptic seizure recognition. Int. J. Intell. Syst. Appl. (IJISA) 5(6), 41–46 (2013). https://doi.org/10.5815/ijisa.2013.06.05 6. Goshvarpour, A., Goshvarpour, A.: Nonlinear analysis of human gait signals. Int. J. Inf. Eng. Electron. Bus. 4(2), 15–21 (2012). https://doi.org/10.5815/ijieeb.2012.02.03 7. Goshvarpour, A., Goshvarpour, A.: Chaotic behavior of heart rate signals during Chi and Kundalini meditation. Int. J. Image Graph. Signal Process. 4(2), 23–29 (2012). https://doi. org/10.5815/ijigsp.2012.02.04 8. Norros, I.: A storage model with self-similar input. Queueing Syst. 16(3–4), 387–396 (1994). https://doi.org/10.1007/BF01158964 9. Kilpi, J., Norros, I.: Testing the Gaussian approximation of aggregate traffic. In: Proceedings of the Second ACM SIGCOMM Workshop, Marseille, France, pp. 49–61 (2002) 10. Sheluhin, O.I., Smolskiy, S.M., Osin, A.V.: Similar processes in telecommunication. Wiley, Hoboken (2007) 11. Ageev, D.V.: Parametric synthesis of multiservice telecommunication systems in the transmission of group traffic with the effect of self-similarity. Electron. Sci. Specialized Ed.: Prob. Telecommun. 1(10), 46–65 (2013) 12. Pashko, A.: Simulation of telecommunication traffic using statistical models of fractional Brownian motion. In: 4th International Scientific-Practical Conference Problems of Infocommunications. Science and Technology. Conference Proceedings, 10–13 October 2017, pp. 414–418 (2017) 13. Pashko, A., Tretynyk, V.: Methods of the statistical simulation of the self-similar traffic. In: Zhengbing, H., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2018. AISC, vol. 754, pp. 54–63. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-91008-6_6 14. Shi, F., Li, C., Zhu, Y.: Robustness evaluation for military communication effectiveness based on multiple data sources and Monte Carlo simulation. Int. J. Mod. Educ. Comput. Sci. 3(5), 1–9 (2011). https://doi.org/10.5815/ijmecs.2011.05.01 15. Mehrdoust, F., Fathi, K., Saber, N.: Accelerated simulation scheme for solving financial problems. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 6(4), 43–48 (2014). https://doi.org/10. 5815/ijitcs.2014.04.05 16. Burkhardt, J.: Bayesian parameter inference of explosive yields using Markov chain Monte Carlo techniques. Int. J. Math. Sci. Comput. (IJMSC) 6(2), 1–17 (2020). https://doi.org/10. 5815/ijmsc.2020.02.01 17. Kozachenko, Y., Pashko, A.A.: Accuracy of simulation of the Gaussian random processes with continuous spectrum. Comput. Model. New Technol. 18(3), 7–12 (2014) 18. Kurchenko, O.O.: One strong consistency estimate of the Hurst parameter of the fractional Brownian motion. Theory Probab. Meth. Stat. 67, 97–106 (2003) 19. Czachórski, T., Doma´nska, J., Pagano, M.: On stochastic models of Internet traffic. In: Dudin, A., Nazarov, A., Yakupov, R. (eds.) ITMM 2015. CCIS, vol. 564, pp. 289–303. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-25861-4_25 20. Kozachenko, Yu., Pogorilyak, O., Rozora, I., Tegza, A.: Simulation of Stochastic Processes with Given Accuracy and Reliability. ISTE Press & Elsevier. London (2016) 21. Pashko, A.O., Rozora, I.V.: Accuracy of simulation for the network traffic in the form of fractional Brownian motion. In: 2018 14th International Conference on Advanced Trends in Radioelecrtronics, Telecommunications and Computer Engineering (TCSET), Slavske, pp. 840–845 (2018). https://doi.org/10.1109/TCSET.2018.8336328 22. Pashko, A., Oleshko, T., Syniavska, O.: Estimation of hurst parameter for self-similar traffic. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2020. AISC, vol. 1247, pp. 181– 191. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-55506-1_16

Research of the Influence of Compromise Probability in Secure Based Traffic Engineering Model in SDN Oleksandr Lemeshko1 , Zhengbing Hu2 , Anastasiia Shapovalova1 Oleksandra Yeremenko1(B) , and Maryna Yevdokymenko1

,

1 Kharkiv National University of Radio Electronics, Nauky Avenue 14, Kharkiv 61166, Ukraine

{oleksandr.lemeshko.ua,oleksandra.yeremenko.ua, maryna.yevdokymenko}@ieee.org, [email protected] 2 National Aviation University, Liubomyra Huzara Avenue 1, Kyiv 03058, Ukraine

Abstract. The paper is devoted to the research of the influence the probability of compromise in Secure Based Traffic Engineering Model in software-defined networks. The flow-based mathematical model under investigation is a further development of the classical Traffic Engineering model. The novelty of the secure based model is the modification of the load balancing conditions for obtaining the routing solution in the software-defined network data plane, which considers not only the network topology, traffic characteristics, and bandwidth of communication links, but also the probability of compromise the network links. As a result, the obtained routing solutions reduced the load on communication links, which have a high probability of compromise, by redirecting traffic to more secure ones. Comparison of Traffic Engineering and secure based Traffic Engineering solutions with different variants of functional dependence of weighting coefficients on the probability of compromise has been performed. The numerical example demonstrates the efficiency of the advanced secure based load balancing model and the adequacy of the obtained order of routing. Keywords: Traffic engineering · Software-defined network · Network security · Probability of compromise

1 Introduction The basic requirements for modern infocommunication networks (ICN) include providing the required level of Quality of Service (QoS), as well as network security [1–4]. At the same time, technological means that are responsible for ensuring QoS and network security often have to trade-off with a negative impact on the efficiency of network operation [1]. For example, the implementation of VPN solutions can reduce network performance, while modern automatic traffic management protocols can compromise the entire level of network security [2]. Therefore, scientific research related to the development and practical implementation of consistent solutions aimed at a controlled improvement of both QoS and network security indicators is relevant. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 47–55, 2021. https://doi.org/10.1007/978-3-030-80472-5_5

48

O. Lemeshko et al.

According to conducted analysis [5–17], the routing means are the most effective mechanisms for QoS and network security improvement. It is the QoS routing protocols as a classic traffic management tool actively expand their functionality towards the proactive and reactive provision of a given level of network security. It should be noted that the most promising among existing is the approach related to the implementation of Traffic Engineering principles [8–17]. In such solutions, the minimum of the upper bound of the network link utilization is selected as the optimality criterion for the realization of the load balancing while transmitting the user flows in the network. Recently, the use of a concept of softwarized networks (SDN, SD-WAN, Hybrid SDN) is a promising towards network development that allows organizing the effective operation of network functions and services together with the simplification of infrastructure and network management [17–19]. Therefore, the scientific task of adapting routing solutions for Traffic Engineering to the requirements of network security by developing appropriate mathematical models that can be used as a foundation of corresponding routing protocols seems to be relevant.

2 Basic Model of Traffic Engineering In this section, the notation that is used in the problem definition is explained. A network   (ICN) is given by a directed graph G = (R, E), where R = Ri ; i = 1, m is the set nodes   and E = Ei,j ; i, j = 1, m; i = j is the set of directed edges representing links between the nodes (network routers). For each edge Ei,j ∈ E (network link) its bandwidth ϕi,j is defined. It should be noted that the number of edges (links) is determined as capacity of this set |E| = n. Let us suppose that in the presented basic Traffic Engineering model, each transmitted flow is unicast with a set of corresponding functional characteristics: • • • •

K – set of transmitted flows of packets (k ∈ K); sk – source router; dk – destination router; λk – k th flow average intensity (packets per second, 1/s).

In order to achieve the solution of the problem of Traffic Engineering, the routing k need to be calculated. Each of the variables correspond to the portion of variables xi,j the k th flow intensity in the link Ei,j ∈ E that is the part of the path. The following constraints imposed on the variables when a single path routing is used [15, 16] k ∈ {0; 1}, xi,j

(1)

k 0 ≤ xi,j ≤ 1.

(2)

as well as for a multipath routing

Research of the Influence of Compromise Probability in Secure

49

The flow conservation conditions are introduced in the model with the aim of ensuring the routes connectivity [16]: ⎧   k k − xi,j xj,i = 0; k ∈ K, Ri = sk , dk ; ⎪ ⎪ ⎪ j:Ei,j ∈E ⎪ j:Ej,i ∈E ⎪  k ⎨  k xi,j − xj,i = 1; k ∈ K, Ri = sk ; (3) j:Ei,j ∈E j:Ej,i ∈E ⎪ ⎪   ⎪ k k ⎪ xi,j − xj,i = −1; k ∈ K, Ri = dk . ⎪ ⎩ j:Ei,j ∈E

j:Ej,i ∈E

Additionally, the average intensity of packets of the k th flow within the link Ei,j ∈ E can be calculated as follows k λki,j = λk xi,j , Ei,j ∈ E,

(4)

To estimate the values of the coefficient of link utilization, the next expression will be used  k k λ xi,j k∈K αi,j = . (5) ϕi,j As shown by the analysis [15, 16], to fulfill the requirements of the Traffic Engineering concept, it is necessary to ensure a balanced use of the available network link resource. This, as a rule, is implemented at the level of formulating conditions for preventing congestion in the network, which in the known solutions have the following form and content: k λk xi,j ≤ αϕi,j , Ei,j ∈ E. (6) k∈K

where α is the additional control variable that numerically determines the upper bound of the network links utilization (5) and obeys the following conditions [18, 19]: 0 ≤ α ≤ 1.

(7)

Finally, as the optimality criterion in the problem solving of the Traffic Engineering technological task the minimum of the boundary value α is selected as it was introduced in [16]: min α. x,α

(8)

The formulation of the Traffic Engineering problem in optimization form with criterion (8) and constraints (1)–(3), (6), and (7) is aimed at ensuring optimal load balancing with minimization of the coefficient of utilization of each of the network communication links. This helps to improve network performance, time QoS indicators (average packet delay and jitter), and reliability indicators such as the probability of packet loss.

50

O. Lemeshko et al.

3 Conditions of Secure Based Traffic Engineering During load balancing in the network, to take into account the indicators of network security, it is proposed to introduce the following modifications into model (1)–(8). Assume that each link Ei,j ∈ E is associated with such an important indicator of network security as the probability of it being compromised pi,j . Their values are known and depend, for example, on the performance of the Intrusion Detection and/or Prevention System (IDS/IPS) installed on the routers. The main goal of the proposed solution is to ensure maximum utilization of communication links with minimal probabilities of compromise, and vice versa – links with a high probability of compromise pi,j should be loaded minimally or even completely blocked. Therefore, it is proposed to use an improved load balancing condition (6): k λk xi,j ≤ αvi,j ϕi,j , Ei,j ∈ E (9) k∈K

where vi,j are the weighting coefficients, which in turn must comply with the following boundary conditions

0, if pi,j = 1; (10) vi,j = 1, if pi,j = 0. If the probability of compromise pi,j increases from 0 to 1, the weighting coefficient vi,j should decrease from 1 to 0. In this paper, as an example, we will consider several variants of the functional representation of the dependence v = f (p) that meet the conditions (10). The simplest case of realization of such function is linear dependence: vi,j = 1 − pi,j ,

(11)

where the utilization factor of the corresponding communication link is directly determined by the level of its security, which is determined by the probability of its compromise. In addition to model (11) in this work for consideration and comparison, an exponential model has been adopted: vi,j = (1 − pi,j )c ,

(12)

where using the coefficient c that takes positive integer values from 1 and above, you can adjust the sensitivity of the network link utilization to values of its probability of compromise. The larger the value of the coefficient c (Fig. 1), the stronger the security parameter affects the blocking of the use of this link. Thus, the secure based Traffic Engineering in SDN task is formulated as an optimization one. Condition (8) is the criterion of optimality, and expressions (1)–(3), (7), and (9) are constraints. When implementing single path routing (1), this optimization problem belongs to the class of mixed linear programming (MILP), and in the case of using multipath routing, it belongs to the class of linear programming (LP).

Research of the Influence of Compromise Probability in Secure

51

Fig. 1. Models of functional dependence of weighting coefficients on the probability of compromise in secure based Traffic Engineering.

4 Numerical Example and Investigation of the Secure Based Traffic Engineering Model on the SDN Data Plane The analysis of the proposed Secure Based Traffic Engineering (SB-TE) model was carried out on a variety of network configurations for a different number of flows and their characteristics. The features of the SB-TE model are demonstrated in the following example. The structure of the network under investigation is shown in Fig. 2, and in the gaps of network links their capacities (numerator) and the probability of compromise R1

R2 800 0.01

---

R4

R3 500 0.03

--R5

400 0.02

--R7

R6 300 0.03

--R8

400 0.01

--R10

R9 500 0.02

--R11

700 0.03

---

R12 600 0.01

---

Fig. 2. Investigated fragment of the SDN data plane.

52

O. Lemeshko et al.

(denominator) are indicated. Within the example, the probabilities of compromising network communication links varied from 0.01 to 0.03. Let the network need to provide a solution to the problem of Secure Based Traffic Engineering for two flows. In this case, both flows are transferred from node R1 to node R12 . Suppose that the characteristics of the transmitted flows vary within the following value: • 1st flow intensity λ1 = 500 1/s; • 2nd flow intensity λ2 = 300 1/s. Within the example, two routing solutions were compared. The first solution (Traffic Engineering) was based on solving an optimization problem with criterion (8) and constraints (1)–(3), (6), and (7). In the solutions of the proposed Secure Based Traffic Engineering model, which were based on solving an optimization problem with an optimality criterion (8) and constraints (1)–(3), (7), and (9) under (11) and (12). The results of the comparison are shown in Table 1. In Tables 1 and 2, the rows are highlighted in gray color, which corresponds to communication links, utilization of which directly determined the value of the optimality criterion (8). Therefore, for the first solution (Traffic Engineering) the upper bound for link utilization was 0.615. It is the links E1,2 , E2,3 , E5,6 , E6,9 , E7,8 , and E9,12 that had the following indicator values (5). For the second solution (SB-TE), the upper bound for link utilization, weighted in relation to the probability of compromise (9)–(11), was 0.632. This value corresponded to the utilization of link E1,2 (5), which was 0.626. For the third solution (SB-TE), the upper bound for link utilization, weighted in relation to the probability of compromise (9), (10), and (12) under c = 2 was 0.649. This value corresponded to the utilization of link E1,2 (5), which was equal to 0.636. For the fourth solution (SB-TE), the upper bound for link utilization, weighted in relation to the probability of compromise (9), (10), and (12) under c = 6, was 0.721. This value corresponded to the utilization of link E1,2 (5), which was 0.679. Analysis of the obtained simulation results presented in Tables 1 and 2 showed that the introduction of coefficients (10) allows in the process of load balancing to reduce the network links utilization in accordance with their security parameters, such as the probability of compromise. This is clearly seen in the example of the change in the link E2,3 utilization with p2,3 = 0.03, when the classical TE model provided the value of the parameter (5) at 0.615, and the use of a linear model of security parameters (11) reduced its value to 0.613. The application of the polynomial model (12) at c = 2 determined the value α2,3 already at the level of 0.611, and at c = 6 this link utilization decreased to 0.6. A similar situation was typical for other communication links with a high probability of compromise such as E3,6 , E6,9 , and E5,6 . However, reducing the overload of the most vulnerable communication links has led to a redistribution of load on the network and increased a load of more secure links, as network bandwidth has remained unchanged. Therefore, within the framework of the proposed secure based Traffic Engineering model in SDN, load balancing considers not only the network topology (6), traffic characteristics, and bandwidth of communication links, but also such important network security indicator as the probability of compromises the network elements.

Research of the Influence of Compromise Probability in Secure

53

Table 1. Comparison of Traffic Engineering and secure based Traffic Engineering solutions – model (11).

Link E1,2 E2,3 E1,4 E2,5 E3,6 E5,4 E5,6 E4,7 E5,8 E6,9 E7,8 E9,8 E7,10 E8,11 E9,12 E10,11 E11,12

pi, j 0.01 0.03 0.02 0.01 0.03 0.02 0.03 0.01 0.01 0.03 0.01 0.02 0.02 0.02 0.02 0.03 0.01

Traffic Engineering solution 1 i, j

2 i, j

492.31 307.69 7.69 184.62 307.69 0 184.62 7.69 0 492.31 7.69 0 0 7.69 492.31 0 7.69

0 0 300 0 0 0 0 300 0 0 238.46 0 61.54 238.46 0 61.54 300

i, j

0.615 0.615 0.385 0.205 0.44 0 0.615 0.44 0 0.615 0.615 0 0.123 0.274 0.615 0.088 0.513

Secure Based Traffic Engineering solution 1 i, j

490.36 306.48 9.64 183.88 306.48 0 183.88 9.64 0 490.36 9.64 0 0 9.64 490.36 0 9.64

2 i, j

10.11 0 289.89 10.11 0 0 0 289.89 10.11 0 0 0 289.89 10.11 0 289.89 300

i, j

0.626 0.613 0.374 0.216 0.438 0 0.613 0.428 0.02 0.613 0.024 0 0.58 0.022 0.613 0.414 0.516

Table 2. Comparison of secure based Traffic Engineering solutions for model (12).

Link E1,2 E2,3 E1,4 E2,5 E3,6 E5,4 E5,6 E4,7 E5,8 E6,9 E7,8 E9,8 E7,10 E8,11 E9,12 E10,11 E11,12

pi, j 0.01 0.03 0.02 0.01 0.03 0.02 0.03 0.01 0.01 0.03 0.01 0.02 0.02 0.02 0.02 0.03 0.01

Secure Based Traffic Engineering Secure Based Traffic Engineering solution (12), c=2 solution (12), c=6 1 i, j

2 i, j

488.41 305.26 11.59 183.15 305.26 0 183.15 11.59 0 488.41 11.59 0 0 11.59 488.41 0 11.59

20.35 0 279.65 20.35 0 0 0 279.65 20.35 0 0 0 279.65 20.35 0 279.65 300

i, j

0.636 0.611 0.364 0.226 0.436 0 0.611 0.416 0.041 0.611 0.029 0 0.559 0.035 0.611 0.4 0.519

1 i, j

480.58 300.36 19.42 180.22 300.36 0 180.22 19.42 0 480.58 19.42 0 0 19.42 480.58 0 19.42

2 i, j

i, j

62.6 0 237.4 62.6 0 0 0 237.4 62.6 0 0 0 237.4 62.6 0 237.4 300

0.679 0.601 0.321 0.27 0.429 0 0.601 0.367 0.125 0.601 0.049 0 0.475 0.091 0.601 0.339 0.532

54

O. Lemeshko et al.

5 Conclusion The paper presents the research of the influence the probability of compromise in secure based Traffic Engineering model in software-defined networks. The proposed mathematical model (1)–(5), (7)–(12) belongs to the class of flow-based solutions and is a further development of the classical model of Traffic Engineering. Within the framework of the model, the secure based Traffic Engineering technological task in a software-defined network data plane is formulated as an optimization problem. The criterion of optimality is the condition (8), and the constraints are expressions (1)–(3), (7), and (9). This optimization problem when implementing single path routing (1) belongs to the class of mixed linear programming (MILP), and when using multipath routing belongs to the class of linear programming (LP). The novelty of the proposed model is the modification of the load balancing conditions in the software-defined network (9)–(12), which in addition to bandwidth (service quality indicator) also considers the network links probability of compromise (network security indicator). In this paper, several variants of the functional representation of the dependence of weighting coefficients on the probability of compromise in secure based Traffic Engineering have been considered (Fig. 1). Therefore, several polynomial models have been proposed and investigated to reckon the influence of network security indicators (probability of compromise) on the order of load balancing in the network. The obtained routing solutions aimed at reducing the load on links, which have a high probability of compromise, by redirecting traffic to more secure ones. The presented numerical example was shown on the SDN data plane. The investigation example (Fig. 2) demonstrates the efficiency of the model and the adequacy of the obtained routing solutions (Tables 1 and 2). All things considered, the future work will be concerned with the ensuring of required values of network security parameters.

References 1. Gupta, S.: Security and QoS in Wireless Sensor Networks. 1st edn. eBooks2go Inc. (2018) 2. Kiser, Q.: Computer Networking and Cybersecurity: A Guide to Understanding Communications Systems, Internet Connections, and Network Security Along with Protection from Hacking and Cyber Security Threats. Kindle Edition (2020) 3. Revathi, S., Geetha, A.: A survey of applications and security issues in software defined networking. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 9(3), 21–28 (2017). https://doi.org/ 10.5815/ijcnis.2017.03.03 4. Lemeshko, O.V., Yevseyeva, O.Y., Garkusha, S.V.: A tensor model of multipath routing based on multiple QoS metrics. In: 2013 International Siberian Conference on Control and Communications (SIBCON) Proceedings, pp. 1–4. IEEE (2013). https://doi.org/10.1109/SIBCON. 2013.6693645 5. Yeremenko, O., Lemeshko, O., Persikov, A.: Secure routing in reliable networks: proactive and reactive approach. In: Shakhovska, N., Stepashko, V. (eds.) CSIT 2017. AISC, vol. 689, pp. 631–655. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-70581-1_44 6. Shaik, M.S., Mira, F.: A comprehensive mechanism of MANET network layer based security attack prevention. Int. J. Wirel. Microw. Technol. (IJWMT) 10(1), 38–47 (2020). https://doi. org/10.5815/ijwmt.2020.01.04

Research of the Influence of Compromise Probability in Secure

55

7. Shashi, R.K., Siddesh, G.K.: QoS oriented cross-synch routing protocol for event driven, mission-critical communication over MANET: Q-CSRPM. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 10(11), 18–30 (2018). https://doi.org/10.5815/ijcnis.2018.11.03 8. Palani, U., Amuthavalli, G., Alamelumangai, V.: Secure and load-balanced routing protocol in wireless sensor network or disaster management. IET Inf. Secur. 14(5), 513–520 (2020). https://doi.org/10.1049/iet-ifs.2018.5057 9. Patil, M.V., Jadhav, V.: Secure, reliable and load balanced routing protocols for multihop wireless networks. In: 2017 International Conference on Intelligent Computing and Control (I2C2) Proceedings, pp. 1–6. IEEE (2017). https://doi.org/10.1109/I2C2.2017.8321936 10. Kumar, N., Singh, Y.: Trust and packet load balancing based secure opportunistic routing protocol for WSN. In: 2017 4th International Conference on Signal Processing, Computing and Control (ISPCC) Proceedings, pp. 463–467. IEEE (2017). https://doi.org/10.1109/ISPCC. 2017.8269723 11. Li, S., Zhao, S., Wang, X., Zhang, K., Li, L.: Adaptive and secure load-balancing routing protocol for service-oriented wireless sensor networks. IEEE Syst. J. 8(3), 858–867 (2013). https://doi.org/10.1109/JSYST.2013.2260626 12. Lemeshko, O., Yeremenko, O., Yevdokymenko, M., Shapovalova, A., Hailan, A.M., Mersni, A.: Cyber resilience approach based on traffic engineering fast reroute with policing. In: 2019 10th IEEE International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS) Proceedings, vol. 1, pp. 117–122. IEEE (2019). https://doi.org/10.1109/IDAACS.2019.8924294 13. Lin, S.C., Wang, P., Luo, M.: Control traffic balancing in software defined networks. Comput. Netw. 106, 260–271 (2016) 14. Lemeshko, O., Yeremenko, O.: Enhanced method of fast re-routing with load balancing in software-defined networks. J. Electr. Eng. 68(6), 444–454 (2017). https://doi.org/10.1515/ jee-2017-0079 15. Lemeshko, O., Yeremenko, O.: Linear optimization model of MPLS Traffic Engineering Fast ReRoute for link, node, and bandwidth protection. In: 2018 14th International Conference on Advanced Trends in Radioelecrtronics, Telecommunications and Computer Engineering (TCSET) Proceedings, pp. 1009–1013. IEEE (2018). https://doi.org/10.1109/TCSET.2018. 8336365 16. Mendiola, A., Astorga, J., Jacob, E., Higuero, M.: A survey on the contributions of softwaredefined networking to traffic engineering. IEEE Commun. Surv. Tutor. 19(2), 918–953 (2017). https://doi.org/10.1109/COMST.2016.2633579 17. Blokdyk, G.: Software-Defined WAN SD-WAN A Clear and Concise Reference. 5STARCooks (2018) 18. Kellerer, W., Kalmbach, P., Blenk, A., Basta, A., Reisslein, M., Schmid, S.: Adaptable and data-driven softwarized networks: review, opportunities, and challenges. Proc. IEEE 107(4), 711–731 (2019). https://doi.org/10.1109/JPROC.2019.2895553 19. Abdullah, M.Z., Al-awad, N.A., Hussein, F.W.: Evaluating and comparing the performance of using multiple controllers in software defined networks. Int. J. Mod. Educ. Comput. Sci. 11(8), 27–34 (2019). https://doi.org/10.5815/ijmecs.2019.08.03

Model of Search and Analysis of Heterogeneous User Data to Improve the Web Projects Functioning Solomiia Fedushko(B) , Oleg Mastykash, Yuriy Syerov, and Anna Shilinh Lviv Polytechnic National University, Lviv, Ukraine {solomiia.s.fedushko,yurii.o.sierov,anna.y.shilinh}@lpnu.ua

Abstract. This study deal with the actual problem for modern science, which is associated with the management of web services, namely the operation management of web projects. The authors developed a model for searching and analyzing heterogeneous user data to improve the functioning of web projects. Means of detection and analysis of heterogeneous data in web project environments are modeled in the work. In this article was analyzed the content of a number of webproject platforms. The authors have developed a general algorithm for searching data in a web project environment and an algorithm for searching heterogeneous data. The collected data from various sources was filtered and consolidated into a single information resource, access to which is provided by the developed data search systems. The proposed methods have improved and simplified the processes of monitoring company employees. Practical implementation has been carried out on popular web projects on Facebook, Instagram, Youtube, TikTok, Pinterest, and LinkedIn. Keywords: Web project · Data search · Model · Analysis · Heterogeneous data · User · Social network

1 Instruction In the most popular web environments, the monthly audience is more than 500 million active users, more than 200 billion web page views per month, more than 1 billion messages are transmitted every day, more than 100 million search queries are performed daily, tens of gigabytes of traffic are generated every hour [1]. According to Cisco [2], it is projected that by 2022 the Internet will have 4.2 billion users, 28.5 billion devices will be connected to the Internet, and 82% of Internet traffic will be transmitted in the form of multimedia content. According to a study [3], the average person spends about 2.5 h a day on traffic in social environments on the Internet, leaving there an extremely large amount of information. According to analytical data, in 2019 the Internet audience numbered 4.39 billion users [4]. Compared to 2014, the number of Internet users in the world has increased by more than 1.9 billion, or 75% in 5 years. Of these, 3.48 billion users were registered in Internet communication environments as of 2019. The most visited sites in the world, YouTube and Facebook occupy the second and third positions, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 56–74, 2021. https://doi.org/10.1007/978-3-030-80472-5_6

Model of Search and Analysis of Heterogeneous User Data

57

and the tenth place is Instagram. Facebook users generate 4 petabytes of data per day according to research in the ranking. The main problem that will be solved is the identification and analysis of heterogeneous data in social media Internet – web projects. The aim is to study arrays of heterogeneous data users located in diverse web projects through the development of new and improvement of existing methods and tools for detection and analysis of user heterogeneous data in web projects. The major research objectives of this work are: 1). analysis of the social environment of the Internet as a source of heterogeneous data and modern approaches and research to identify user data in web projects; 2). construction of models of the environment of user data and information content in web projects; 3). construction of methods and algorithms for detection, structuring and storage of heterogeneous data in diverse web projects. Despite the large number of different solutions for data discovery and analysis in web projects, today the task of building methods of adaptive analysis of heterogeneous interconnected data in web projects, in a constantly changing data and their structure, remains unsolved. All available web project analysis techniques only partially solve the problem of collecting information from such web environments. This technique of gathering information, as manual analysis pages though there are universal-term, but has a major drawback - a very low processing speed and da are large costs of human resources. Existing approaches to automated data collection in web projects allow you to quickly obtain data, but existing automated data collection systems have a number of disadvantages. For example, the data are partial and highly specialized, usually the data structured in the form convenient for marketing purposes, the data do not provide a full understanding portrait of friendly environments and such data are often unreliable, is not able to track user data historicity change the social environment of the Internet. The experience of operating a number of systems of analysis of heterogeneous data of web projects presented on the market has shown that they can effectively solve the problem of identifying the most popular topics discussed in web projects by specified parameters, as well as display detailed user information. important for analytical work. Meanwhile, these systems have significant drawbacks. The main one is the incompleteness of information collection. In addition, the services broadcast a large amount of information “garbage”, important information is lost in the queue of insignificant messages. Thus, the scientific task of developing methods for analyzing the detection and analysis of heterogeneous user data in web projects is that there are no universal methods for collecting heterogeneous, dynamic and interconnected data of any web project with their subsequent consolidation into a single information resource. Although to date there is a large amount of research on the analysis of web projects [5–7].

58

S. Fedushko et al.

2 Related Works The formation of an understanding of the web project environment was studied by M. Kilduff, W. Tsai, D. Hansen, B. Shneiderman, M. A. Smith [8] and others. The researchers presented web projects as a graph of interconnected nodes, where the vertices of the graph are user profiles with their attributes, and the edges are the links between them. The current area of research are analysis of social networks and user content of their users [9, 10], technologies for data analysis [11–14], heterogeneity of data in social networks [15, 16], information technology [17], verification and analysis of data of users of web projects [18], management of virtual communication environments [19, 21]. A web project is a platform whose main purpose is to provide functionality to ensure the communication of its users. The result of communication of web project users [11–14] is the generation of heterogeneous data. Heterogeneous data analysis is one of the areas of data analysis that has been studied for many years. In particular, scientists have built models of integration of heterogeneous data [22], taking into account the peculiarities of information resources of web-systems; the definition of the basic principles and ways of distribution of processes of integration of values of heterogeneous data, their syntax, structure and semantics is investigated. The first step in the study of data heterogeneity is to highlight the concept of a web project in which such data is generated. The concept of a web project was discovered in the studies of Lodhia S., Stone G., Peleshchyshyn A. [8], Huminsky R., Wu Y., Outley C., Matarita-Cascante D. and others. However, the developed methods for managing virtual projects are imperfect. Therefore, the model of search and analysis of heterogeneous user data to improve the functioning of web projects is an urgent task.

3 Methodology of the Analysis of Heterogeneous User Data to improve the Web Projects Functioning 3.1 Development of a Common Data Search Algorithm in a Web Project Environment Search algorithm requires a user-entered input data set, which will later be used to form criteria and algorithms for data retrieval in the initial stages. The input data set is formed according to the constructed user data model. The model of the input data set is shown in Fig. 1. The input data contains two types of features: • basicFirstName; LastName; MiddleName; Email; PhoneNumber; Photo; Video; Voice. • additional: Place of study; Work place; Hobby; List of probable locations with possible indication of time periods; Friends list; Interests and interests of the user; Auxiliary multimedia data; Additional Information.

Model of Search and Analysis of Heterogeneous User Data

59

Primary signs are signs by which the user can be uniquely identified and characterized. The primary features are fixed, strictly defined and cannot be added by the user. during the process of working with the software-algorithmic complex. As a rule, they are unique and do not occur in other users. Secondary features are features that do not allow you to identify the user, but complement the primary features. These features allow you to specify the main features and give a more detailed description of the user’s portrait. Typically, they can be repeated by different users. Secondary features can be entered manually by the user, specifying the attributes that must be used to search for these features. If the parameter is subsequently successfully identified and found, it will be added to the input dataset.

Input Model - FirstName : String - LastName : String - Email : String - PhoneNumber : String - Photos: List - Audios : List - Videos : List - Friends: List - Addresses: List

Main Features

Features

Fig. 1. Input data set model

One of the prerequisites for using the proposed method is the presence of a registered user. The procedure of registration of a new user in the system is an important component of the system, which allows to uniquely identify the user, give him the functionality to view statistics of the system, start the process of analysis and verification of data entered and build a system of recommendations according to his profile. The user registration procedure consists of the following steps: 1. 2. 3. 4. 5. 6.

user authorization data entry; validation of entered data; checking the presence in the user’s system; sending a verification code to the user; confirmation of registration; search for information about the registered user in existing web projects and check the entered authorization data.

Mandatory authorization data include the following data: Login; Password; Name; Email address or phone number; Preliminary criteria for data search. The step of supplementing the data includes filling in additional information of the system user about himself (Personal photo; Gender; Year of birth; Place of work / study; Current geolocation).

60

S. Fedushko et al.

The new user registration algorithm is shown in Fig. 2.

User

Analyzer

Entering authorizaon data

Checking the correctness of the entered data

MessageController

SearchModule

[Incorrectly entered data]

Checking user availability

[User is already registered]

Sending verificaon code

Confirmaon of registraon

Data addion Search for user data in SEI

Fig. 2. Algorithm for registering a new user on the web-project platform

Searching for registered user data on existing web project platforms is the last step. This search is performed by a separate service running in the background. The search results do not affect the registration procedure, but are necessary to identify the user and collect his personal information. The data obtained as a result of the search is stored in the database for use as part of the database. 3.2 General Algorithm for Searching Heterogeneous Data An authorized user has the ability to search for data in a web project. The general algorithm for searching for heterogeneous data in web project environments is shown in Fig. 3.

Model of Search and Analysis of Heterogeneous User Data

61

Fig. 3. General algorithm for searching data in a web project environment

The user selects and fills in the required search criteria manually or selects from the template, followed by its completion. Information is entered through the UI application (Android, Windows). Available search options: • Specialized search. Specify the specific community (part of the community) that will be searched. • Widely specialized search. Search all available communities with a combination of search results. • Approximate search. If previous searches did not return results, you will be able to select this type of search. In all cases, the user has the ability to flexibly configure the criteria (attributes) by which the search is performed. An information entry template is an entity that contains pre-selected nodes (criteria) of an input data set with partial or complete content to search for users of the. platform of all projects. The template is created by the administrator and each registered program

62

S. Fedushko et al.

participant. The template created by the administrator is available to all participants. The template created by the user is available only to this user. The resulting set of input data must be analyzed for correctness of input, divided into structurally related portions of information, which are used to search for data in social environments of the Internet. The process of data splitting and validation is described in the algorithm for detecting heterogeneous data in web project environments. Finding existing user data in the database is the next step. This step allows you to speed up the search for the required data. When receiving pre-saved data, the following criteria are taken into account: • • • •

Date and time of the last update of a specific node in the database; Completeness of stored data; Data connectivity; Data matching.

The generated data criteria affect the stage of data collection and analysis. In particular, they have an impact on the following processes: • • • • •

the need to re-search the data; the need to re-analyze the data; historicity and frequency of data change; forming a profile portrait of the desired user; reduce the load on background process servers.

The transition to the stage of selecting sources (web project platforms) for which the data is searched, occurs after the formation of the input data set. The program administrator adds information search sources. 3.3 Algorithm for Adding a New Data Source The process of adding the following search source is not completely automatic. The actions that the user must perform at this stage are as follows: • enter the domain name of the web project; • enter the user data on whose behalf the program will run. Enter multiple users for reliable software operation; • enter the required headers for web project requests. This step is optional and is performed only if incomplete authorization data (entered in the previous step) is entered; • introduction of additional information for program training. This data depends on the characteristics of the web project. The program has a set of appropriate rights to access community data. For correct operation of the program the service of appeals of the registered user to a platform is configured. First, the program is authorized to successfully obtain the source document of the user’s page. The authorization procedure for the user looks like this:

Model of Search and Analysis of Heterogeneous User Data

63

• go to the authorization page; • enter login and password; • fill in other required data in the appropriate fields (captcha, secret key). The algorithm for adding a new source is shown in Fig. 4.

User

Data Service

Validator

Outside service

Enter a domain name Request Pinging

return

return

return

Create a Request

return

return

return

Cycle Sending headlines

return

return Add metadata to the query

Data storage

Fig. 4. Sequence diagram for adding a new data source

After successful authorization, the user is redirected to a personal page in a web project environment. The procedure for authorizing a platform in a web project is slightly different: • first you need to get the url address of the authorization page;

64

S. Fedushko et al.

• generate a request with a login, password and other necessary data; • send a request (almost always a POST request) to this page and wait for a response. If the answer contains information about the successful authorization, then the token and other necessary data for further operation of the program are saved. The authorization request contains: • • • • •

URL of the page; Headlines; set-cookie; parameters that are passed as part of the navigation bar or in the body of the request; additional parameters (ajax, DOM-tree data).

The session opens after the user is authorized on the server of the web project platform. The response from the server includes the following data: • parameters: server variables on the basis of which the DOM-tree of the page is formed; • set-cookie; • status response code. Typically, the session ID is stored in cookies or in the DOM-tree (depending on how the web project environment works). The user’s token is transmitted in cookie files or in one of the response parameters. The token, session ID and other required parameters (browser version) are checked by the server for each subsequent request to identify the user. If the request fails, an error code is returned from the server. A token is a digital signature of a user’s login and password. In different technologies token generation differs in algorithm and principle, for example: bearer token, access_token, token, custom_token and so on. But they have one thing in common: the pair, for example (Fig. 5):

Fig. 5. Server request headers

The token lifetime is determined on the server side of the web project environment. It is most often deleted when deleting a user session. The token is one of the main keys to accessing project data, but not the only one. In addition to this parameter, other header keys are also used. When adding a new resource, the administrator must enter them manually. After entering this data, the web project server checks the correctness of the input by sending a request to the target resource, if the answer is correct (for

Model of Search and Analysis of Heterogeneous User Data

65

example, the status of the response code will be 200), this resource is added to the list of available sources. The program also automatically generates headers based on the received answer/answers. 3.4 Algorithm for Obtaining Information on Web Pages of Web Project Environment Profiles After the process of registering the platform and selecting resources for data retrieval, the next stage of data collection takes place. In Fig. 6 presents a universal algorithm for obtaining information of web project platform profiles.

UI module

Finder

module

DAL module

Checking the correctness of the entered data and user login

data

Return the required data for the user from the database

Data output search for data

Request submission

Get the search page for the search pages Request a specific page Analysis of the obtained data and

Data storage

format There is another search page

Stage

Data output

Fig. 6. Algorithm for obtaining information on web pages of a web project platform

The first step in this stage is to form the structure of web pages. If in the web project environment there is a user search functionality according to the specified criteria, then a query to the search page is formed, which contains data intended to refine the search and we send a query to the social environment of the Internet to search for the user. We

66

S. Fedushko et al.

store links to pages of found profiles. If there is no search functionality, then we use recursive search of the user’s web page in the web project environment. The initial vector of the process is the user who searches for data. A separate module for searching for a mention of the user in the web project environment is also launched in parallel. The algorithm for forming the structure of web pages to search for information is shown in Fig. 7.

SearchModule

ServiceHelper

HttpClient module

DAL module

Read from the database of pre-saved settings

Step 1

Get SEI settings

There is a search functionality

Generate and send a user search query

Described in more detail in another algorithm

Getting a response from SEI Preparation of queries for recursive search and search for user references

Answer analysis

Save the answer

Recursive search

Answer analysis

Save the answer

Search for mentions Answer analysis

Save the answer

Stage

Output of work results

Fig. 7. Algorithm for obtaining the structure of web pages for information retrieval

If the user data and profile are listed in the database of the web project, then a report on the available information with the date. In case of receiving updated user data, a request is made for all personal pages of the web project, which are stored in the database. The information retrieval process is performed again. Finding a new user in a web project environment:

Model of Search and Analysis of Heterogeneous User Data

67

• Go to the user search page in the web project platform (the address is stored in the database). Fill in the necessary parameters to search, send a request, get an answer. • If user profile information is found, it is saved in the database on the web project platform. • If no results are found, the user is searched for throughout the web project environment and the resource address (posts, comments, photos) is saved. There can be three possible answers: 1) the user’s web page was found, 2) the page was not found, but there are mentions of the user, 3) the user information was not found. The next step is to read the required information from the web pages of the web project form. After receiving the user’s profile address, a request is made to the web project environment to obtain the DOM tree of the page. In some cases, when obtaining a complete page tree of one query is not enough for analysis, it is necessary to re-send asynchronous queries to the web page with a change of input parameters (number of read web page, ID of the next query, etc.). The rule for reading the tree is set at the stage of adding the web project environment to the list of sources. The next step in obtaining data is to validate this data: • check for correctness of syntactic formation of the document. If there are incorrectly formed elements, then either we remove them or try to bring them to the correct form (a separate module does it). Depends on the importance of the element (we will have a dictionary of nodes that are critical to us - to paste it somewhere above and in the database); • check the correctness of links to other pages (broken links); • delete links to other platforms that we do not need; • search and delete broken links to resources; • removal of scripts, auxiliary elements, etc. 3.5 Recursive User Search There is a high probability of re-entering the analyzed web page database when recursively browsing the pages of network users. In order to avoid duplication of web pages in the database, an associative collection is created, which stores the IDs of already passed users. If the user ID is already stored in the collection, the system does not analyze the page of such user. The algorithm of recursive user search in the web project environment is shown in Fig. 8. Recursive data retrieval is slow and resource-intensive, but it allows you to create a structured and hierarchical tree of interconnected nodes of the web project environment (Fig. 9). The stage of checking the correctness of such data, which provides a better analysis of the content, is used as an additional stage of filtering the data. The sequence diagram for data validation is shown in Fig. 10. According to the sequence diagram for checking the correctness of the data, the data is checked in several stages (Fig. 10):

68

S. Fedushko et al.

Fig. 8. Recursive user search in a web project environment

• Parsing. Selection of grammatical connections between words. This stage is implemented by a separate processing module, which filters options that do not constitute a semantic load and are illogical. Each entered character is checked separately, for example, the entered e-mail and phone number are checked by a regular expression. • Semantic analysis. Determining the desired meaning of each word. Thus in knowledge base each word is attached to it by a certain value depending on values of surrounding words.

Model of Search and Analysis of Heterogeneous User Data

69

User Name Photo Birth Date Work place List of educaonal instuons Friends / communies list List of posts

Fig. 9. Providing data with a certain format

Data Service

Grammar Module

Module

Call

Return Call

Return

Fig. 10. Sequence diagram for data validation

4 Experiments One of the important results of this work is the modeling of the process of searching and analyzing heterogeneous user data to improve the functioning of web projects. This model was tested by analyzing web projects implemented on popular platforms, namely: Facebook, Youtube, Instagram, LinkedIn, Pinterest, TikTok.

70

S. Fedushko et al.

The following approaches were used for testing: search for user data in different environments of web projects with their subsequent consolidation into a single profile for each user; quantitative analysis of users of each web project environment.

Fig. 11. Results of age analysis of web project users

Period of active research of web project management system: September 2020 November 2020. During this period, a total of 959485 user profiles of web projects and 7684 web projects were researched and analyzed. The studied audience of users is divided into groups according to age and sex of users. The results of age analysis are presented in Fig. 11. Users are divided into groups according to the following age characteristics: • • • • •

from 15 years to 24 years; from 25 years to 35 years; from 36 years to 47 years; from 48 to 59 years; from 60 years and above.

Model of Search and Analysis of Heterogeneous User Data

71

The analysis of activity of researched groups of users is also carried out. Quantitative indicators of user activity were the posts they published on personal web pages or on other pages of web projects. The results of analysis are presented in Fig. 12 and 13.

Fig. 12. Statistics of publishing posts by users of web projects

The user of the web project environment performs actions on posts. These are reactions to the post. Quantitative indicators of reactions to posts are presented in Fig. 13.

Fig. 13. Response to published posts by users of web projects

Another indicator used to study web project environments is the frequency of visits to web project environments by users, namely daily visits to web projects. The results are shown in Fig. 14. As can be seen from the results of all graphs, the activity of the age audience of users depends on the specific environment of web projects, their direction and specialization, but it is worth highlighting the age category of users from 18 to 24 years who are the most active users of web projects.

72

S. Fedushko et al.

Fig. 14. Statistics of daily visits of web projects by users.

5 Conclusion The paper modeled the means of detection and analysis of heterogeneous data in web project environments. The content of a number of web projects platforms of higher educational institutions of Ukraine, other state institutions and official pages of representatives of these institutions was also analyzed. The information collected from various

Model of Search and Analysis of Heterogeneous User Data

73

sources is filtered and consolidated into a single information resource, access to which is provided by developed data search systems. Another important result of the work is the development of a database that consolidates information from diverse web projects. The real-time database is updated and filled with additional data through the operation of the developed background modules. System administrators are responsible for the correct filling of the database, which, together with automatic analyzers, makes it possible to keep real user data up to date. The results of the proposed methods are useful for firms to monitor potential job candidates. The proposed methods have improved and simplified the processes of monitoring company employees. Also, the results of the services can be used to find the most suitable applicants for admission to the year of study. The main criteria for selecting candidates: area of interest, places of previous training, profile of previous training, posts and distribution, circle of friends and their interests, interest in the department, communication in the field of projects. Practical implementation and experimental research was conducted on popular web projects on Facebook, Instagram, YouTube, TikTok, Pinterest and LinkedIn. Potential consumers of the developed methods in this study of detecting and analyzing heterogeneous user data in web projects are organizations and specialists involved in the data analysis system, such as marketers, PR-specialists of companies, political parties, celebrities and security services. Acknowledgment. This research is supported by National Research Foundation of Ukraine within the project “Methods of managing the web community in terms of psychological, social and economic influences on society during the COVID-19 pandemic”, grant number 94/01-2020.

References 1. Digital 2019: Global Internet Use Accelerates [wearesocial]. – Mode of access. https://wea resocial.com/blog/2019/01/digital-019-global-internet-use-accelerates 2. VNI Global Fixed and Mobile Internet Traffic Forecasts (2019). https://www.cihttps://www. cisco.com/c/en/us/solutions/service-provider/visual-networking-index-vni/index.htmlsco. com/c/en/us/solutions/service-provider/visual-networking-index-vni/index.html 3. Newman, N.: Reuters institute digital news report 2019. Reuters Institute for the Study of Journalism (2019) 4. Castano, S., De Antonellis, V.: Global viewing of heterogeneous data sources. IEEE Trans. Knowl. Data Eng. 13(2), 277–297 (2001) 5. Bhatti, S., Dewani, A., Maqbool, S., Memon, M.: A web based approach for teaching and learning programming concepts at middle school level. Int. J. Mod. Educ. Comput. Sci. (IJMECS) 11(4), 46–53 (2019). https://doi.org/10.5815/ijmecs.2019.04.06 6. Maulidiah, S., Sitanggang, I., Sukoco, H.: ICT training recommendation using web usage mining. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 10(12), 21–26 (2018). https://doi.org/10. 5815/ijitcs.2018.12.03 7. Sharma, K., Marjit, U., Biswas, U.: RDF Link Generation by exploring related links on the web of data. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 10(10), 62–68 (2018). https://doi.org/ 10.5815/ijitcs.2018.10.08

74

S. Fedushko et al.

8. Peleshchyshyn, A., Mastykash, O.: Analysis of the methods of data collection on social networks. In: International Scientific and Technical Conference “Computer Science and Information Technologies”, Lviv, Ukraine, 05–08 September (2017) 9. Fedushko, S., Ustyianovych, T., Gregus, M.: Real-time high-load infrastructure transaction status output prediction using operational intelligence and big data technologies. Electronics 9(4), 668 (2020). https://doi.org/10.3390/electronics9040668 10. Hryshchuk, R., Molodetska, K., Syerov, Y.: Method of improving the information security of virtual communities in social networking services. In: CEUR Workshop Proceedings. 1st International Workshop on Control, Optimisation and Analytical Processing of Social Networks, COAPSN-2019, vol. 2392, pp. 23–41 (2019) 11. Khan, P.W., Byun, Y., Park, N.: A data verification system for CCTV surveillance cameras using blockchain technology in smart cities. Electronics 9(3), 484 (2020) 12. Gregus, M., Kryvinska N.: Service orientation of enterprises - aspects, dimensions, technologies, Comenius University in Bratislava (2015). ISBN 9788022339780 13. Molnár, E., Molnár, R., Kryvinska, N., Greguš, M.: Web Intelligence in practice. J. Serv. Sci. Res. 6(1), 149–172 (2014). https://doi.org/10.1007/s12927-014-0006-4. ISSN 2093-0720 14. Poniszewska-Maranda, A., Matusiak, R., Kryvinska, N., Yasar, A.-U.-H.: A real-time service system in the cloud. J. Ambient. Intell. Humaniz. Comput. 11(3), 961–977 (2019). https:// doi.org/10.1007/s12652-019-01203-7. ISSN 1868-5137 15. Wang, L.: Heterogeneous data and big data analytics. Autom. Control Inf. Sci. 3(1), 8–15 (2017) 16. Thomas, J., Sael, L.: Overview of integrative analysis methods for heterogeneous data. In: International Conference on Big Data and Smart Computing, pp. 266–270 (2015) 17. Kysil, T., Izonin, I., Hovorushchenko, O.: Information technology for choosing the trademark considering the attitude of consumer. In: CEUR Workshop Proceedings, vol. 2623, pp. 133– 140 (2020) 18. Shi, C.: A survey of heterogeneous information network analysis. IEEE Trans. Knowl. Data Eng. 29(1), 17–37 (2016) 19. Baako, I., Umar, S., Gidisu, P.: Privacy and security concerns in electronic commerce websites in Ghana: a survey study. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 11(10), 19–25 (2019). https://doi.org/10.5815/ijcnis.2019.10.03 20. Abdur, R., Belal, H., Sharifur, R., Saeed, S.: Subset matching based selection and ranking (SMSR) of web services. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 11(4), 44–53 (2019). https://doi.org/10.5815/ijitcs.2019.04.05 21. Ojugo, A.A., Eboka, A.O.: Assessing users satisfaction and experience on academic websites: a case of selected nigerian universities websites. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 10(10), 53–61 (2018). https://doi.org/10.5815/ijitcs.2018.10.07 22. Berko, A.Yu., Alekseeva, K.A.: Processing of inhomogeneous data in information resources of WEB-systems. Bulletin of the Lviv Polytechnic National University. Information systems and networks. № 814, pp. 23–32 (2015)

Modeling the Dynamics of “Knowledge Potentials” of Agents Including the Stakeholder Requests Andrii Bomba1 , Taras Lechachenko2(B) , and Maria Nazaruk3 1 National University of Water and Environmental Engineering, Rivne, Ukraine

[email protected] 2 Ternopil Ivan Puluj National Technical University Ternopil, Ternopil, Ukraine

[email protected] 3 Rivne State Humanitarian University, Rivne, Ukraine

Abstract. According to expert estimates, professionals who can study throughout life, think critically, set goals and achieve them, work in a team, communicate in a multicultural environment, and possess other modern knowledge and skills will be the most successful in the labor market in the near future. That is why, in this article, the authors analyze the processes of educational and qualitative personality growth based on diffusion-like models of the process of knowledge potential dissemination and redistribution. In particular, an emphasis is put on the process description (modeling) of the knowledge potential redistribution in the formation of the system of professional competencies. The conceptual and nonlinear mathematical models of information processes of the agents’ “knowledge potentials” formation are developed taking into account their constituent components and customer requests. In particular, the procedure of constructing the previously proposed multi-component two-dimensional array of discrete values of the knowledge potential constituent components for generating procedures to enhance (improve) the professional competencies of knowledge sources, is generalized and specified. The general approaches to the formation of the corresponding optimization problems are determined (in particular, the minimization of the employer and educational institution costs and the time for training an agent in the dual system). Keywords: Professional competencies · Agent · Knowledge potential · Dual education · Stakeholder · Marketing environment · Educational institution

1 Introduction In a changing and closely interconnected world, everyone needs a wide range of skills and competences that need to be developed throughout their lives [1]. Therefore, one of the contemporary visions is to analyze and study the formation of a person-oriented educational trajectory that is a personal educational (training) program of a person, which provides with the acquisition of professional competences that correspond to © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 75–88, 2021. https://doi.org/10.1007/978-3-030-80472-5_7

76

A. Bomba et al.

their abilities, interests, motivation, psychodynamic and emotional characteristics, age and level of knowledge potential (for example, in the form of hard skills and soft skills, etc.). The formation of a person-oriented educational trajectory of an individual should take place on an adaptive basis with a balance between the requests of customer and the personal abilities of the agent. Solving this dilemma is the key to effectively developing professional competencies that correspond the needs of the market environment actual state. Formalization of the knowledge dissemination process including the customer requests (in dual education), building a model of the process is an urgent task, since the integration of the industrial and educational sectors depends on the effectiveness of modeling knowledge potentials in this system. The innovative development of the economy depends on the effectiveness of stakeholders collaboration in the dual education system. The need to build a model of knowledge dissemination including the customer requests is due to the complexity of the integration of stakeholders in the dual education system (company and educational institution), taking into account the dominance of various parties depending on the country of implementation of this form of education, the complexity of their cooperation. Thus, the construction and analysis of mathematical and information models of the formation of teachable and measurable abilities (hard skills) that can solve complex problems, is a necessary and urgent task. In this study, the task is to formalize the model of spreading knowledge potential, including customer requests (in the system of dual education).

2 Related Literature Review An analysis of studies [2, 3] have showed that a main of scientists attention is paid to develop models that reflect the interaction processes between the objects of the educational segment and the labor market and the modernization of educational programs, as well as the education system in general. At the same time, these results cannot be directly applied in practice in all cases. In the study [4] authors examine the transform of professional knowledge between explicit knowledge and tacit knowledge, the expansion model of SECI is introduced into the exploratory analysis. In paper [5] researchers provides an understanding of factors that involved in implementing knowledge management concept to enhance organizational performance. In particular, in paper [6] authors examine the proposed solutions of the general educational problem. An agent-based modeling approach to study the knowledge spreading was used by Morone & Taylor, Xuan, Hirshman, Giacchi and others. These studies mainly focused the influence of network (group of agents) properties on knowledge spreading within that network (organization). In particular, Morone & Taylor, using a cellular vending machine model, studied how knowledge is carried in a network in which agents interact through face-to-face communication. Their scientific works performs the effect of network structure on the learning process [7–9]. In the paper [10], present an improved knowledge diffusion hypernetwork (IKDH) model based on the idea that knowledge will spread from the target node to all its

Modeling the Dynamics of “Knowledge Potentials” of Agents

77

neighbors in terms of the hyperedge and knowledge stock. In this paper, a model of knowledge dissemination is considered, which is characterized by network parameters and agents’ stock of knowledge, that is, research is focused on the properties of the network and agents, the speed of knowledge transfer. In the study [11], authors propose a knowledge transmission model by considering the self-learning mechanism and derive the mean-field equations that describe the dynamics of the knowledge transmission process. Based on the fact that the people prefer to spread knowledge to their acquaintances, a friendship-based altruistic incentive knowledge diffusion model (SIH model) is proposed in the paper [12]. This study also examines the process of knowledge dissemination in various types of networks using the example of students from different courses. In the model proposed in study [9], agents (nodes) transfer knowledge based on two criteria, namely knowledge distance and confidence. They show that in a more distributed network configuration, the dynamics of knowledge discussion and the conviction level are much greater than the centralized structure [13]. In general, in the studied knowledge spreading models, knowledge is represented in a few ways as a stock (each agent has an initial knowledge level, which is an integer selected randomly from a given range), as a tree where a little potential knowledge conforms to each node or as a set of facts that characterizes the agent’s knowledge matrix [14]. The studies under consideration examine the process of knowledge dissemination at a certain level of abstraction. There are no examples of knowledge spreading research in specific areas, in particular in specific educational systems. For example, the system of dual education requires separate consideration, since the formation of knowledge potential in this system occurs in two different environments (cliques). The effectiveness of the functioning of this system is determined by the degree of integration of the constituent subsystems, cooperation of stakeholders [15].The dual education system reflects the interaction of the educational sector with the industry of the country of operation. The relationship in the growth of innovation from the knowledge diffusion between education institution and industry is explored in the works [16, 17].Given the importance of the effective functioning of the dual education system, the study of the knowledge potential dissemination requires special attention in this area. The existing studies do not examine the adaptive characteristics of educational objects and there are no models of knowledge transfer between subjects of different educational environments, so there is an attempt to describe the information processes of knowledge distribution in the form of a generalized diffusion model. As result, the purpose of this work is to analyze and develop a mathematical model of the redistribution of potential of knowledge processes in the derivation of the agent’s professional competence system including the customer requests.

78

A. Bomba et al.

3 Main Part 3.1 The Problem Formulation The numerical characteristic that helps fix a some level of agent’s knowledge, gathered during training, life experience, etc., has been suggested in our previous works [19, 20] to call as knowledge potential ϕ. It has been constructed the so-called diffusion-liked models of information processes of knowledge potential redistribution in the educational social and communication environment of the city and focuses on the description (simulation) of processes of knowledge potential redistribution in the system formation of professional competences in particular, taking into account the influence of customer (companies). The competence usually means a functional combination of intellection, ways of cogitative, views, evaluate, skills, capacity, and other private properties of that determines a person’s capability to successfully pursue professional and further educational activities [20]. It is estimated that a particular competence is described by a definite of components of knowledge potential and their constituents. Using φq,l,k,m,ε , it is denoted the value of q constituent of l knowledge potential component of k agent at time (l = 1, l∗ , q = 1, q∗ , k = 1, k∗ ). For example, Fig. 1 presents some characteristics of competence formation based on the constituents and components of the knowledge potential including the customer requests.

Fig. 1. Components of competencies and the environments of their acquisition

In Fig. 1 present the formation of competencies in the system of the dual type of education (where the customer is the company in which the student is studying simultaneously with the educational institution). The competencies that are formed in the environment of the educational institution are transformed and supplemented through expansion in the environment of the company, taking into account its requests.

Modeling the Dynamics of “Knowledge Potentials” of Agents

79

Training in the company represents the adaptive side of the knowledge potential formation at the request of the actual development of the labor market, without displacing the academic component of training. The simultaneous of knowledge potential formation in two cliques (company and educational institution) of the dual form of education minimizes the lag in updating educational programs of an educational institution through the effective integration and adaptation of learning in two environments. In general, as in [20], the set of all possible components of the knowledge potential constituents at a given time situation can be characterized by a matrix (two-dimensional array (with the size l × q)): ⎞ ϕ1,1,k,m , ϕ2,1,k,m , · · · ϕq∗ ,1,k,m ⎜ ϕ1,2,k,m , ϕ2,2,k,m , · · · ϕq ,1,k,m ⎟ ∗ ⎟ ⎜ ⎠ ⎝... ... ... ... ϕ1,l∗ ,k,m , ϕ2,l∗ ,k,m , . . . ϕq∗ ,l∗ ,k,m ⎛

The complex of potential subsets of such an array will be joined with the set of competences gained by agents during educational [19, 20]. It is shown some examples of such subsets on Fig. 2, that compose competencies. In Fig. 2 present the transformation of the competencies expansion at m + 1 instant of time. To simplify the display, the disturbing parameters E are not indicated and the designation k of the agent is taken such that assigned is by default.

Fig. 2. Schematic representation of professional competence expand

80

A. Bomba et al.

The figure illustrates that at each next m + 1 moment of time m1 , m2 , m3 , the competencies formed from the components of the knowledge potential are expanded and supplemented with new components. It should be noted that the nature of acquiring knowledge potential, in particular, the formation of competencies occurs through the combination of knowledge potential components, which may be common for different competencies (i.e., overlap). The subsets of competencies increase with each next step, including as a result all the elements of the set of knowledge potential components. In (Fig. 2) present the complexity of the subsets formation of the knowledge potential components A, B, C due to their union and overlapping by figures of arbitrary shape. In Fig. 3 it is shown the process of forming competencies from the point of view of reflecting their expansion in the dynamics of a particular specialty studied in an educational institution. At each moment of time, the process of increasing subsets of competencies and adding new ones to those already existing at m + 1 moment in time is schematically indicated. In the figure present the process of obtaining competences in an educational institution during the period of study, their expansion and the emergence of new ones at certain fixed stages (course) of education.

Fig. 3. Dynamics of competence formation in m + 1 moment of time

It is set out to form such subsets to satisfy certain optimal conditions for training costs and program outcomes, with regard to the needs dictated by stakeholder conditions. That is, taking into account customer requests in the dual education system based on the interaction of the training parties through the effective integration of the educational process in two environments. 3.2 Mathematical Model of the Customer Requests Process The acquisition of professional competences and the implementation of program learning outcomes in model of the customer requests are the result of four forms of interaction [21–23]:

Modeling the Dynamics of “Knowledge Potentials” of Agents

81

Agent / Knowledge Source (teacher / lecturer); Agent / Agent; Agent / Educational Content Agent/Company. In this paper, it is described not only the interaction between agents (learners), knowledge sources and learning content, but also the feedback between teachers and agents, company (instructors), as well as the educational content used to enhance the level of knowledge potential (Fig. 4) at the request of the customer. The increase in the level of knowledge potential in the dual education system occurs through the feedback of the teacher and the instructor of the company with the educational content that the agent (student) obtains.

Fig. 4. Forms of interaction between agents, teachers, company and educational content

The effectiveness of an agent training in two cliques depends on the degree of training integration in the educational institution and the company. This synergy is achieved by the feedback of the agent with the teacher and the instructor in the company and the interaction between the teacher and the instructor mediated by the common transformation of the educational content. The mathematical model of training in dual education is determined by the parameters, which are set the proportional distribution of obtaining the knowledge potential in two cliques (in educational institution and company). The effectiveness of this training depends on the consistency of feedback connections between the components and the stakeholders of the system. Thus, modeling the dynamics of knowledge potentials at the request of customers is determined by the following parameters:

82

A. Bomba et al.

distribution of the obtaining of knowledge potential in two cliques, feedbacks between agents, agents and educational content. As mentioned in [19], the process of knowledge potential redistribution (including the acquired knowledge) at the next time stage without considering the knowledge redistribution between agents is represented as: ϕq,l,k,m+1 = ϕq,l,k,m + αq,l,k,m+1 f∗ q,l,k,m+1 +

(1)

∗ , 0 < αq,l,k,m+1 < 1 (1 − αq,l,k,m+1 )fq,l,k,m+1

where f∗ q,l,k,m+1 is the main part of the characteristic of the knowledge source ∗ is the part of the characteristic of the knowledge source cus(teacher), and fq,l,k,m+1 tomer (instructor in company), αq,l,k,m+1 , and (1 − αq,l,k,m+1 ) corresponding weights, which, in particular, characterize the share of knowledge acquired by a certain agent (pupil, student, etc.) in a particular year in the corresponding departments (curricula, educational programs of educational institutions, as well as the corresponding plans of customer companies, etc.). This construction, in particular, allows to adapt the academic component of training to the requirements and tasks that the student must solve at a certain (prospective) workplace (using in practice the knowledge gained in an educational institution). Thus, training, for example, of a company in a dual form of education adapts and complements the acquisition of competence in an educational institution, primarily a practical component. The acquired competencies reflect the complex of knowledge and skills acquired in practice at the request and requirements of the customer’s company (certainly, that f = αf ∗ + (1 − α)f∗ - is a common characteristic of two types of sources of knowledge, in a certain case characterizes the essence of duality). We can also directly (more simplified) form the time redistribution of the constituent components of knowledge potentials, for example, as follows: q , l , k ,1

q , l , k ,0

f q , l , k ,1

f

q , l , k ,0

q , l , k ,1 * q , l , k ,1

(1

q ,l , k ,1

)f

,...,

* q ,l , k ,1

(2) m 1

q ,l , k , m 1

(

q , l , k ,0 m 1

q ,l , k , m 1

f

* q ,l , k , m

(1

*

q ,l , k ,1

) f q ,l , k , m ),

∗ where, ϕq,l,k,0 = ϕq,l,k,0 - a discrete function characterizing the initial level of knowledge of the k-th agent, and the indices q and l run through the given sets of values (for example, according to Fig. 2). In previous works, cases of modeling the process of formation of “final” knowledge (taking into account the symmetrical difference between the acquired knowledge of the educational institution and the needs of the enterprise) were considered. Taking into account the “final” components of competencies, the parameters αq,l,k,m˜ will be selected optimally depending on the form of feedback from agents (regarding the professional qualification of knowledge sources) goals and program learning outcomes that characterize the required professional competencies and optimal costs. In the simplest case, the following conditions can be formed in the form of inequalities:

ϕ q,l,k,m+1 < ϕq,l,k,m+1 < ϕ q,l,k,m+1 ,

(3)

Modeling the Dynamics of “Knowledge Potentials” of Agents

83

or (in expanded form) ϕ q,l,k,m+1 < ϕq,l,k,0 +

m+1 

(αq,l,k,m+1 f∗ q,l,k,m˜ +

m=1 ˜

∗ (1 − αq,l,k,1 )fq,l,k, m ˜ ) < ϕ q,l,k,m+1 ,

(4)

where φ q,l,k,m+1 , φ q,l,k,m+1 - certain threshold values of the components of knowledge potential and their constituents, which must satisfy the program results of the corresponding specialty. The disadvantage of the proposed model is that directly by formulas (2), (3), (4),… we do not “see” the competencies acquired directly in the education institution, firm, together. So in the future we will represent the knowledge potential by the sum.    + ϕq,l,k,m + ϕq,l,k,m ϕq,l,k,m = ϕq,l,k,m

(5)

where the strokes characterize the knowledge potentials obtained, respectively, in an educational institution, firm, together. In Fig. 5 schematically present the “final” situation, where such characteristics are depicted in green, blue and blue-green, respectively, red indicates the competencies required by the company on request. Let’s indicated the costs of training a specialist from the educational institution and   , βq,l,k,m . the company, respectively βq,l,k,m We offer several options for the design of the objective function, present the optimization problem of minimizing the cost of agent’ (student) training at the request of the customer (in the system of dual education) as follows: m+1 

 (βq,l,k,m αq,l,k,m˜ f∗ q,l,k,m˜

m=1 ˜

 ∗ + βq,l,k,m (1 − αq,l,k,m˜ )fq,l,k, m ˜ ) → min .

(6)

Taking into account also the expenses for the student (in particular the scholarship), we generalize formula 6 as follows: m+1  m=1 ˜

  ∗  (βq,l,k, ϕ ˜ f∗ q,l,k,m ˜ + βq,l,k, ˜ )fq,l,k,m m αq,l,k,m m (1 − αq,l,k,m m ˜ + γq,l,k, m q,l,k,

  + γq,l,k, ϕ  m + γq,l,k, ϕ  m ), m q,l,k, m q,l,k,

(7)

   where γq,l,k, , γq,l,k,m ,γq,l,k, corresponding weights coefficient. m ˜ m ˜ If the essence of cost optimization lies in minimizing the agent preparation time from the part of the customer in a dual system, then it is necessary to optimize m. Considering the limitations at corresponding intervals:

(0.40k∗ E − 0.75k∗ E), (0.25k ∗ E − 0.60k ∗ E),

(8)

84

A. Bomba et al.

Fig. 5. Example of “Final” situational state of knowledge potentials illustration, which takes into account all types of acquired knowledge, where the circles show “miss” (missed components of knowledge, insufficient assimilation, etc.)

where E is the ECTS credit, k∗ is the number of study credits which contains the discipline q dedicated for training an agent in the dual education system (according to the Law of Ukraine, training in the dual education system between the parties (the company and the educational institution) is distributed in the range from 25% to 60% in a company and in an educational institution from 40% to 75% respectively). Note that, for example, a firm must satisfy the inequality: m+1 

 βq,l,k,m αq,l,k,m˜ f∗ q,l,k,m˜ ≤ F,

(9)

m=1 ˜

where F is the company’s fund designed for personnel training costs. Solving the relevant tasks allows us to take into account the customer’s requests in the process of accumulating knowledge potential, makes it possible for the customer to influence the final result of the agent’s training (namely optimization of the task provides the ability to adapt the academic component of training, while minimizing the customer’s costs and thereby maximizing the benefit from direct participation in agent’s training). As a result of solving the optimizing problem certain functionals (for example, optimizing costs), in conditions, for example (4), we find the parameters on the basis of which, according to formulas (1–3), we take into account the characteristics of the corresponding “perturbation” of educational programs to take into account, for example, the needs dictated by the conditions stakeholders. Remark. Similarly [19, 20], we also take into account the diffusional redistribution of knowledge potentials due to communication between agents and knowledge acquired as a result of the additional sources of information impact (television, libraries, private lessons, etc.): ∗ ϕq,l,k,m+1 = ϕq,l,k,m + αq,l,k,m+1 f∗ q,l,k,m+1 + (1 − αq,l,k,m+1 )fq,l,k,m+1

Modeling the Dynamics of “Knowledge Potentials” of Agents ∗

+

k 

δq,l,k, k,m (ϕq,l, k,m − ϕq,l,k,m ) +

k=1

k∗  k=1

∗ δq,l,k, f∗ , k,m+1 ∗,q,l,k, k,m+1

85

(10)

where k ∗ , k∗ - the number of agents of the clique and additional sources of infor∗ , - coefficients that characterize the effect of mation, respectively, δq,l,k, k,m and δq,l,k, k,m knowledge transfer in communication.

4 Discussion On Table 1 present the modeling of the growth dynamics of the average value of the agent’s knowledge potential during 4 years of training. The first line of Table 1 reflects the dynamics of changes in the average value of the agent’s knowledge potential level with the generation of initial values of the knowledge potential level of instructors, teachers, and the agent’s coefficient of perception. The second row of the table shows the dynamics of change of the average of the agent’s knowledge potential level at the condition of the generation of high intervals of the knowledge potential level of instructors (in the company) at average values of the agent’s coefficient of perception and the average level of knowledge potential of teachers. The third line shows the values at the condition of programming of the high level of the teachers potential and the average values of the instructors potential level (in the company), the average values of the agent’s coefficient of perception. The fourth row of the table reflects the agent’s knowledge potential at the condition of the high level of teachers and instructors knowledge potential with the average level of the agent’s perception. The last row of the table present the agent’s knowledge potentials at the condition of generation the high level of agent’s perception at the average values of teachers and instructors potentials. From the data in Table 1 it can be seen that with the same increase in the level of teachers and instructors potentials (rows 2 and 3) the level of agent’s potential is higher at condition of increasing potential of teachers than instructors, this can be explained by a higher percentage of training in educational institution in this simulation. The highest level of the agent’s potential is observed at the condition of high indicators of the coefficient of the agent’s perception and in another case also at high indicators of the level of teachers and instructors potential (at average values of the coefficient of perception of the agent (line 4)). Thus, we can conclude when the agent’s knowledge perception is the average (below average) it is necessary to increase the percentage of training in the environment whose level of sources of knowledge (potential) is higher. It should be noted that at a high level of the coefficient of the agent’s perception - the level of knowledge potential is the highest. This fact can be explained by the high ability of the agent to self-study, natural talent.

86

A. Bomba et al.

Table 1. Generating the dynamics of the average values of the agent’s knowledge potential, taking into account the parameters The value of knowledge potentials

Year of study 1

2

3

4

Values generation in the initial of their levels

64,25

106,40

156,05

205,05

Value at generating a high level of instructors potential (in the company)

69,24

114,61

160,27

205,77

Value at generating a high level of teachers potential

71,73

120,57

168,37

216,35

Value at generating a high level of teachers and 74,41 instructors potential

125,00

175,86

227,59

Value at generating a high coefficient of agent’s perception

158,37

225,43

292,36

90,93

5 Conclusions The paper proposes models of information processes for redistributing the agent’s knowledge potential in the dual education system, taking into account its constituent components in the formation of professional competencies system. In particular, a nonlinear mathematical model of knowledge dissemination including the stakeholder requests was built. The dynamics of the agent’s knowledge potential growth at different levels of knowledge potential of teachers, instructors and taking into account the coefficient of the agent’s perception, as well as the different ratio of training time in the company and the institution was carried out. The results of the study indicate the need for adaptive adjustment of the curriculum in the dual education system, depending on the agent’s ability to learn (the level of the perception coefficient) and taking into account the level of knowledge potential of teachers, mentors. The prospect of further research is the solution to the proposed in the work, optimization problems of the distribution of knowledge potentials in the dual education system, which will allow the customer to influence the final result of the agent’s training (that is, to adapt the academic component of training, while minimizing the customer’s costs and maximizing the benefit from direct participation in agent training). Another task of further research is to build a mathematical model of the consideration of the parties influence proportion (company and educational institution) on evaluation training results of the agent in the dual system.

References 1. Annexto the Proposal for a Council Recommendation on Key Competences for Lifelong Learning. https://ec.europa.eu/education/sites/education/files/annex-recommendationkey-competences-lifelong-learning.pdf

Modeling the Dynamics of “Knowledge Potentials” of Agents

87

2. Lisichk, E., Postnikova, E., Tverdokhlebov, S.: Formation of professional competence of students in engineering education. Creat. Educ. 3, 104 (2012) 3. Lengelle, R., Meijers, F., Poell, R., Geijsel, F., Post, M.: Career writing as a dialogue about work experience: A recipe for luck readiness? Int. J. Educ. Vocat. Guidance 16(1), 29–43 (2015). https://doi.org/10.1007/s10775-014-9283-1 4. Yang, J., Pan, Y.: Analysis on teaching methods of industrial and commercial management based on knowledge transform expansion model of-SECI. IJ Educ. Manag. Eng. 6, 467–473 (2011) 5. Zaied, A.N.H., Hussein, G.S., Hassan, M.M.: The role of knowledge management in enhancing organizational performance. Int. J. Inf. Eng. Electron. Bus. 4(5), 27 (2012) 6. Wibig, T., Dam-o, P.: The model of the evolution of the knowledge content and contemporary science education crisis. Int. J. Mod. Educ. Comput. Sci. 4(1), 61 (2012) 7. Morone, P., Taylor, R.: Knowledge diffusion dynamics and network properties of face-toface interactions. J. Evol. Econ. 14(3), 327–351 (2004). https://doi.org/10.1007/s00191-0040211-2 8. Xuan, Z., Xia, H., Du, Y.: Adjustment of knowledge-connection structure affects the performance of knowledge transfer. Expert Syst. Appl. 38(12), 14935–14944 (2011) 9. Hirshman, R., Charles, J., Carley, K.: Leaving us in tiers: Can homophily be used to generate tiering effects. Comput. Math. Organ. Theory 17, 318–343 (2011). https://doi.org/10.1007/ s10588-011-9088-4 10. Wang, J.P., Guo, Q., Yang, G.Y., Liu, J.G.: Improved knowledge diffusion model based on the collaboration hypernetwork. Phys. A 428, 250–256 (2015) 11. Wang, H., Wang, J., Ding, L., Wei, W.: Knowledge transmission model with consideration of self-learning mechanism in complex networks. Appl. Math. Comput. 304, 83–92 (2017) 12. Zheng, W., Pan, H., Sun, C.: A friendship-based altruistic incentive knowledge diffusion model in social networks. Inf. Sci. 491, 138–150 (2019) 13. Giacchi, E., La Corte, A., Di Pietro, E.: A dynamic and context-aware model of knowledge transfer and learning using a decision making perspective. In: Proceedings of the 1st International Conference on Complex Information Systems, pp. 66–73 (2016) 14. Argote, L., Fahrenkopf, E.: Knowledge transfer in organizations: the roles of members, tasks, tools, and networks. Organ. Behav. Human Decis. Process. 136, 146–159 (2016) 15. Gessler, M.: The lack of collaboration between companies and schools in the German dual apprenticeship system: historical background and recent data. Int. J. Res. Vocat. Educ. Train. (IJRVET) 4(2), 164–195 (2017) 16. Lööf, H., Broström, A.: Does knowledge diffusion between university and industry increase innovativeness? J. Technol. Transf. 33(1), 73–90 (2008). https://doi.org/10.1007/s10961-0069001-3 17. Rupietta, C., Backes-Gellner, U.: How firms’ participation in apprenticeship training fosters knowledge diffusion and innovation. J. Bus. Econ. 89(5), 569–597 (2018). https://doi.org/10. 1007/s11573-018-0924-6 18. Cowan, R., Jonard, N.: Network structure and the discussion of knowledge. J. Econ. Dyn. Control 28(8), 1557–1575 (2004) 19. Bomba, A., Nazaruk, M., Kunanets, N., Pasichnyk, V., Bilak, Y.: Modeling the redistribution processes of knowledge potential in the formation of the professional competency system. In: 14th International Scientific and Technical Conference on Computer Sciences and Information Technologies, Lviv Ukraine, pp 197–201 (2019) 20. Bomba, A., Nazaruk, M., Pasichnyk, V., Kunanets, N.: Modeling the dynamics of knowledge potential of agents in the educational social and communication environment. In: Shakhovska, N., Medykovskyy, M.O. (eds.) Advances in Intelligent Systems and Computing IV, vol. 1080, pp. 17–24. Springer, Heidelberg (2019). https://doi.org/10.1007/978-3-030-33695-0_2

88

A. Bomba et al.

21. Lane, A.: Placing students at the heart of the iron triangle and the interaction equivalence theorem model. J. Interact. Media Educ. 10, 1–8 (2014) 22. Mulder, F.: The Logic of national policies and strategies for open educational resources. Int. Rev. Res. Open Dist. Learn. 14(2), 96–104 (2013) 23. Fujiwara, D.: Valuing the Impact of Adult Learning: An Analysis of the Effect of Adult Learning on Different Domains in Life, p. 28 (2012)

Crop Irrigation Demand Modelling Method Within the Conditions of Climate Change in the Dnipro-Donbas Canal Pavlo Kovalchuk1 , Viktoria Rozhko1 , Volodymyr Kovalchuk1 , Hanna Balykhina2 , and Olena Demchuk3(B) 1 Institute of Water Problems and Land Reclamation NAAS, Kyiv 03022, Ukraine 2 National Academy of Agrarian Sciences of Ukraine, Kyiv 01010, Ukraine 3 National University of Water and Environmental Engineering, Rivne 33028, Ukraine

Abstract. The presence of a tendency to rise in temperature and increase the dryness of the climate requires assessing changes in water demand in irrigation in a particular region. However, the existing indices of climate change do not provide such estimates for the period of vegetation growing crops. The method of estimation of tendencies of relative change of crop water demand during irrigation during the long-term period is offered. The method is based on a comparison of multi-year water availability curves due to the deficit of the water balance of crops during the periods of vegetation. The annual change in water need for agricultural crops for two not overlapping periods is estimated. The model estimates the tendency of crop water demand changes when irrigation carried out in a multi-year section. Model based on the moving average method. The method and model are applied to the assessment of changes in water demand in the area of the DniproDonbas canal, Ukraine. A comparative analysis of the relative change in the crop water demand for grain corn and alfalfa was carried out according to the data of six meteorological stations. The results can be used to assess the prospects of irrigation development in the region. The method and models can be used in water resources management systems based on the basin principle. Regional studies are developing estimates of changes in water demand for irrigation in the context of climate change, which requires the development and use of new approaches, methods and models. Keywords: Crop water demand modeling method · Dnipro-Donbas canal area · Drought · Irrigation rates · Water availability · Water balance deficit

1 Introduction Under the conditions of gradual warming and emergencies humanity has dealt with negative developing process of natural phenomena such as multiplication of the events and intensity of droughts [1]. Presence of a strong tendency towards increasing of the climate aridity and rising of the summer maximum temperatures are proved during the last decade [2]. As a result, agricultural crops need more amount of water during irrigation. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 89–101, 2021. https://doi.org/10.1007/978-3-030-80472-5_8

90

P. Kovalchuk et al.

Many different indices were offered by researchers to characterize aridity and climate change [3–5]. The soil moisture deficit indices [3, 6, 7] and evapotranspiration deficit indices [6] were proposed for agricultural monitoring. In order to highlight areas for different levels of water availability and determine reasonability of growing certain crops, hydrothermal coefficient of wetting (HTC) was proposed by Selyaninov and is used in Ukraine[8–10]. HTC is calculated using a formula: HTC = P * 10/t, where P is precipitation in millimeters within period with temperatures above +10 °C, t is determines the amount of temperature in degrees °C for the same period. A zoning of Ukraineis carried out in terms of hydrothermal resources endow [11]. Algorithms and the basic principles of the aridity level assessing or moisture content of the territory, particularly using obtaining normalized values of the standardized precipitation index (SPI), are given in [4, 12, 13]. SPEI index calculates the difference between P-precipitation and PET-potential evaporation monthly (or weekly) [14]. The index includes the main aspects of the water balance such as water input and water output. In the international practice for regional studies the new approach for modeling of the water balance of the Nile River basin was described [14]. It demonstrates the differences in distribution between rainfall, evapotranspiration and productivity of water within the basin. A similar approach is used for European conditions in [12]. However, these indices do not take into account changes of crop water demand (water needs) for irrigation. Today, there are many measures for minimizing of the negative impact of drought [15], but the most effective one is irrigation reclamation [16]. Our research interests involve the criteria for water needs of agricultural crops within identified are using water balance models. These criteria include irrigation rates. In contrast to the work [17], the goal of predicting the water demand of agricultural crops is not set. Research focuses on the actual changes in water demand in the context of climate change that have occurred in the region over the past 56 years, namely from 1960 to 2015. It is necessary to develop methods for assessing of crop water demand changes when irrigation in a particular area of Ukraine. These areas are not provided with water resources and water supply is carried out using canals. The research offers a new method of evaluating the changes in water demand for crops under irrigation in the conditions of climate change, as in the case of the region of the Dnipro-Donbas canal, Ukraine.

2 The State of the Problem In world practice, system models of water resources management [18, 19] and automated water management systems [20] are being developed. The use of water in certain areas also requires an integrated approach [21]. Such territories are not provided with water resources and water is supplied using canals. One of the important sources of area for water providing is a Dnipro-Donbas canal system (Fig. 1). Water is delivered from the Kamyanske reservoir to the canal for different needs, such as water service for population of Kharkiv city and Kharkiv district, the Siversky Donets River basin environmental rehabilitation as well as the water supply for

Crop Irrigation Demand Modelling Method

91

irrigation systems. Within monitoring system water quality is defined for drinking needs, irrigation, as well as using environmental assessment method [22]. Such estimates based on the use of neural networks [21].

Fig. 1. Sketch map area of Dnipro-Donbas canal system and sampling points

Remote water quality monitoring systems are being developed [23]. In a comprehensive study, it becomes necessary to develop methods for assessing changes in water demand for irrigation. However, the existing approaches to forecasting water demand based on neural networks [24] are not adapted to the conditions of agricultural irrigation. Therefore, a method, in which stochastic curves of crop water supply are used as the basis for calculating the change in water demand, is being developed [25]. A method, that compares these stochastic curves in different periods of time and estimates the total increase in water demand in years with different moisture supply, is proposed. To calculate water demand, data, obtained over many years at meteorological stations in the Dnipro-Donbas Canal area, are used. The data of systematic observations of meteorological parameters are the basis of numerous experiments on the proposed models.

3 Methodology and Models 3.1 Calculation of Water Balance Deficit of Crops within the Territory To analyze the dynamics of water demand use a water availability graph according to a water balance deficit of crops. It was developed by [25]. So the crop water demand is reflected in the water balance deficit. The total deficit for the growing season is a biologically optimal irrigation rate. It is proposed a method, which involves comparing these graphs for different periods as well as assessment of the total growing of water demand for different years. The data of the meteorological stations were used for calculating of water demand within Dnipro-Donbas canal area. Average water balance deficit Di over ten-day period for major crops is calculated according to the data of meteorological stations within the canal area by the formula: Di = Ri − EToi , i = 1, . . . , N ,

(1)

where EToi is evapotranspiration during ten-day period, mm; Ri is total precipitation during ten-day period, mm.

92

P. Kovalchuk et al.

To assess the impact of climate change on the water demand of agricultural crops, it is the water balance deficits that are most characteristic. Taking into account the initial and final reserves of moisture in the soil, the productive factors of the culture is inappropriate to use when solving this problem. FAO recommended using Penmann-Monteith method to determine the total evapotranspiration [26]. The local biophysical method by Shtoyko has been approbated in Ukraine [27]. Comparison of the estimates of evapotranspiration for the mentioned methods for calculating the evapotranspiration is given in [28]. By the Shtoyko’s formula for evapotranspiration is calculated when having incomplete plant shadowing:  n a  (2) tci 0, 1 tci − Ec1 = i=1 100 and in other periods when having complete plant shadowing:  n a  . (3) Ec2 = tci 0, 1 tci + i=1 100  Where ni=1 tci is amount of average daily air temperatures for n days, °C; tci is average daily air temperature, °C; a is average daily degree of relative air humidity, %. In case of partial shading of the soil by plants until complete shading, we recommend to use the formula for the linear dependence of expressions (2) and (3): Ec3 = (1 − ϕ)Ec1 + ϕEc2 , 0 ≤ ϕ ≤ 1,

(4)

where ϕ = 0, if there is no shading, ϕ = 1, if there is full shading. The Shtoiko method allows calculating moisture evaporation by the field Ec, occupied by a certain crop for any period of time and water demand during the growing season. The advantage of this method is the possibility of using observational data on temperature and relative humidity over a long-term period from 1960 to 2015, at meteorological stations in the region of the Dnipro-Donbas canal. For example, for alfalfa of the second and subsequent years in spring in the first 10 days after the renewal of the growing season, the coefficient ϕ varies from 0 to 1; similarly, during 10 days after each slope a linear combination of (2) and (3) is used. In the final phases of the growing season, the need for irrigation already becomes equal to 0. In the area of the Dnipro-Donbas canal formula (2) is applied for corn from sowing (3rd decade of April) to germination (2nd decade of May). For the cover (appearance of 10–14 leaves in the 3rd decade of June), formula (4) is used. Formula (3) is used from the appearance of 10–14 leaves in all active stages to the beginning of the milky-wax ripeness of the grain (2nd decade of August). Formula (4) is applied from the beginning of the milky-wax ripeness of the grain to the maturation of the grain. But watering of corn stops in the phase of milky-wax ripeness of grain. The growing season of alfalfa in the second and subsequent years lasts from April 1 to October 1. Usually three cuts of alfalfa are carried out in the area of the Dnipro-Donbas canal. Based on the restrictions of salt concentration in the irrigated water the irrigation rates are determined based on the quality of each source of irrigation water.

Crop Irrigation Demand Modelling Method

93

Water use rate of agricultural crops ωj for a particular year j are calculated based on the total deficit of water balance of crops during their growing season, taking into account the specificity of modern crop cultivation technologies: ωj =

l i=1

Di , j ∈ [1; N ],

(5)

where Di is deficit of water balance of crops during ten-day period; l is number of ten-day periods over growing season of the plant. It was observed stochastic nature of variation of required cumulative water balance deficits ωj for years (Fig. 2). To form an array of gradual growth of deficits arranging was done taking into account N years. Probability of the year P(ωj ) % was calculated according to the total water balance deficit by the formula [25]: P(ωj ) =

j − 0, 3 ∗ 100%, N + 0, 4

(6)

ωj, m3/ha

where j is ranked number in the series; N is number of years in a row.

Series1, 2007, Series1, 1998, Series1, 1975, 1981, Series1, 2009, Series1, Series1, Series1, 1968,1972, Series1, 1962, Series1, 2013, Series1, 1996, Series1, 1961, 2012, Series1, 1995, Series1, 1965, 4778.603 Series1, 2008, Series1, 1979, 4482.812 2011, Series1, 1967, 2006, Series1, 1963, 4341.752 Series1, 1986, Series1, 1960, 4309.292 4156.609 2002, 2010, Series1, 1966, 4080.644 4033.855 Series1, 2014, Series1, 1983, 4005.266 3949.61 3914.888 3901.84 1999, Series1, 1971, 3837.861 Series1, 1964, 3794.243 3751.074 Series1, 1993, Series1, 2003, Series1, 1970, 3670.182 3627.46 3604.262 Series1, 2000, 3557.303 3558.004 3494.602 2005, 3442.008 Series1, 1994, 3435.293 Series1, 1989, Series1, 1984, Series1, 1992, 3247.626 3234.257 3211.187 Series1, 1987, Series1, 1991, 3155.822 Series1, 2015, Series1, 1982, 3130.936 3008.722 Series1, 1990, 2987.184 2929.073 Series1, 1985, 2865.748 2865.004 2795.718 2707.056 Series1, 1969, Series1, 1988, Series1, 1980, 2582.341 2559.633 2555.341 2488.508 2476.811 Series1, 2004, Series1, 1997, 2323.565 2319.952 Series1,1973, 1974, 2267.432 2260.607 Series1, 2001, 2129.201 Series1, 2042.05 1824.364 1780.906 1726.558 Series1, 1976, 1576.574 1481.847 Series1, 1978, 1429.395 1366.601 1229.587 841.848 560.945 Series1, 1977, -656.872 N, years

Fig. 2. Stochastic graph of total water balance deficit for corn based on data of Hubinyha meteorological station for 1960–2015 period

ω, m3/ha

Using calculated probability of the year (6) a graph of total water balance deficit rising was completed from the wettest year to the driest one (Fig. 3).

Series1, 98.76, Series1, 96.99, 95.21, 93.44, Series1, 91.67, 89.89, 88.12, 86.35, 84.57, 82.80, 81.03, 79.26, 77.48, 75.71, 73.94, 72.16, Series1, 70.39, Series1, 68.62, Series1, 66.84, Series1, 65.07, Series1, 63.30, Series1, 61.52, Series1, 59.75, Series1, 57.98, 4778.603 Series1, 56.21, Series1, 54.43, Series1, 52.66, Series1, 50.89, Series1, 49.11, 4482.812 Series1, 47.34, Series1, Series1, 43.79, 45.57, Series1, 42.02, 4341.752 4309.292 Series1, 40.25, 4156.609 Series1, 38.48, Series1, 36.70, Series1, 34.93, 4080.644 Series1, 33.16, 4033.855 Series1, 31.38, 4005.266 3949.61 3914.888 3901.84 Series1, 29.61, Series1, 27.84, 3837.861 Series1, 26.06, Series1, 24.29, 3794.243 3751.074 Series1, 22.52, 3670.182 3627.46 3604.262 Series1, 20.74, 3558.004 3557.303 3494.602 3442.008 3435.293 Series1, 18.97, 2707.056 Series1, 17.20, Series1, 15.43, 3247.626 3234.257 3211.187 3155.822 3130.936 Series1, 13.65, Series1, 11.88, 3008.722 2987.184 Series1, 10.11, 2929.073 Series1, 8.33, 2865.004 2865.748 2795.718 Series1, 6.56, 2582.341 2559.633 2555.341 2488.508 2476.811 Series1, 4.79, 2323.565 2319.952 2267.432 2260.607 2129.201 Series1, 3.01, 2042.05 1824.364 1780.906 1726.558 1576.574 1481.847 1429.395 1366.601 1229.587 Series1, 1.24, 841.848 560.945 656.872 P(ω), %

Fig. 3. Completed graph of total water balance deficit for corn based on probability of the year using data of Hubinyha meteorological station for 1960–2015 period

94

P. Kovalchuk et al.

3.2 Method for Estimating Change in Water Demand by Crops During study of changes in water use for irrigation due to climate change entire series of observations was divided into two periods: N 1 is the number of years for the first period, N 2 = N − N 1 is number of years for the second period. Based on the data presented [29] there has been a rapid rise in temperatures on Earth since the second half of the 1980s. While from 1960 to 1985 there was no increase in temperature. This was used when selecting two time intervals of the same duration of 28 years each to compare the water demand of crops (Fig. 4).

Fig. 4. Choosing of two representative periods of climatic data for analysis water demand

A ranking of total water balance deficit for each period was done in ascending order for the certain plant: ω1 ≤ ω2 ≤ · · · ≤ ωN1 ,

(7)

ωN1 ≤ ωN1 +1 ≤ · · · ≤ ωN2 .

(8)

ω, m3/ha

The graph demonstrates water availability based on total water balance deficit for each period of the study (Fig. 5).

1960-1987, 1988-2015, 1988-2015, 1988-2015, 1960-1987, 1988-2015, 1988-2015, 1988-2015, 94.01, 97.54, 1988-2015, 1988-2015, 1960-1987, 1988-2015, 1960-1987, 1988-2015, 1960-1987, 1960-1987, 1988-2015, 90.49, 1960-1987, 1988-2015, 86.97, 83.45, 1960-1987, 1988-2015, 1960-1987, 1960-1987, 1988-2015, 79.93, 1988-2015, 76.41, 1960-1987, 1988-2015, 1988-2015, 1960-1987, 1988-2015, 94.01, 90.49, 1960-1987, 72.89, 1960-1987, 1960-1987, 86.97, 1988-2015, 1960-1987, 1988-2015, 1960-1987, 69.37, 1988-2015, 1960-1987, 1960-1987, 65.85, 4419.5655 4559.694 4813.135 62.32, 1960-1987, 1988-2015, 58.80, 76.41, 79.93, 1960-1987, 51.76, 1960-1987, 72.89, 48.24, 69.37, 65.85, 44.72, 4035.931 62.32, 1960-1987, 23.59, 41.20, 3919.7065 3876.151 83.45, 3254.57 37.68, 51.76, 48.24, 34.15, 3682.862 30.63, 3647.6255 44.72, 55.28, 2970.17 27.11, 41.20, 3509.301 3498.3185 37.68, 3417.3715 34.15, 1988-2015, 58.80, 2766.45 3348.0095 20.07, 55.28, 2710.97 16.55, 30.63, 3269.533 13.03, 27.11, 1988-2015, 23.59, 3166.8835 3116.7615 20.07, 9.51, 2980.791 16.55, 2946.1005 2947.2345 2935.9065 13.03, 2905.7025 2886.881 2875.023 2873.9125 2841.9205 2792.753 9.51, 2277.7705 2705.2565 2610.526 2587.2335 2547.374 2474.4745 2422.3675 2362.963 2347.713 2347.8195 2208.127 1960-1987, 2170.0925 5.99, 1540.133 2100.924 2078.1465 2053.0175 2015.9565 2014.382 2.46, 1378.998 1980.516 1960-1987, 1783.2905 1762.8925 1707.4875 1691.1325 1515.8785 5.99, 350.799 2.46, 115.748 1960-1987

1988-2015

P(ω),%

Fig. 5. Graphs of water availability based on total water balance deficit for different periods using data of Izyum meteorological (for corn)

Crop Irrigation Demand Modelling Method

95

Rising trend of irrigation rate for different periods was evaluated according to areas S 1 and S 2 which bounded by integral lines of water availability based on total water balance deficit (Fig. 5):  N1   (9) ωJ dp ωj , S1 =  S2 =

1 N2

  ωJ dp ωj .

(10)

N1

Growth of crop water demand under irrigation due to climate change is calculated using the formula: Is =

S2 − S1 . S1

(11)

It demonstrates the percentage change of water needs for crops in comparison with basic data of observations S 1 . 3.3 Model of Change of Crop Water Demand Under Irrigation Using the Method of Moving Average Estimation model for crop water demand changes under irrigation based on the choice of a basic series of years, corresponding to the initial level of water demand S0, and comparison with water demand S i of estimated series of years within i interval. They are selected according to moving average method [25]. The relative change of crop water demand for each interval from n intervals is calculated as: Isi =

Si − S0 , i = 1, . . . , n. S0

(12)

Moving average method is that the forecasted criteria are calculated as the average value of the criteria in the prior periods. In general, the formula has the form: Si =

1 m ωj , j=1 m

(13)

where S i is the average value of water demand for identify intervals of time; m – number of years within interval, which used for calculating of average value of water demand; ωj is value of total water balance deficit within i is interval of study time. To calculate water needs S i time interval shift is defined by moving average method for m years, which follow after the basic years. During applying of moving average method, the larger interval of averaging for time series deficits, the smoother line of crop water demand changing.

4 Results of Numerical Experiments and Discussion To calculate crop water demand data of meteorological stations within Dnipro-Donbas canal area were used: Kobelyaky (Poltava region), Hubinyha (Dnipropetrovsk region), Krasnograd, Izyum, Lozova and Slobozhansky (Kharkiv region) (Fig. 6).

96

P. Kovalchuk et al.

Fig. 6. Location of meteorological stations within the Dnipro-Donbas canal area

Assessing of crop water demand changes due to climate change, the graphs of water availability within growing season for corn and alfalfa were completed for 1960–1987 period and 1988–2015 one (Fig. 7, 8).

Fig. 7. Graphs of water availability based on total water balance deficit for alfalfa at points of the meteorological stations for two periods

Graphics objective analysis shows growth of crop water demand for 1988–2015 period compare with 1960–1987 one for the selected meteorological stations. Increasing is clearly demonstrated for dry and wet years. Growing season of alfalfa has a long period and amounts seven months from March to September. Corn vegetates during summer from May to September and irrigation

Crop Irrigation Demand Modelling Method

97

Fig. 8. Graphs of availability based on total water balance deficit for corn at points of the meteorological stations for two periods

season involves only three months from middle of May to middle of August. Increasing of crop water demand refer to different durations of the growing season of the plants. Graphics objective analysis shows growth of crop water demand for 1988–2015 period compare with 1960–1987 one for the selected meteorological stations. Increasing is clearly demonstrated for dry and wet years. It was established that relative change of crop water demand during growing season is developing within the Dnipro-Donbas canal area: 13% for alfalfa and 22% for corn. That means peaking of drought phenomena is observed in summer (growing season of corn) (Table 1). To assess crop water demand trends model was used for moving average method when time series of water needs were overlapped. The basic years for calculating of water demand S0 were 1960–1980 period, duration consists 21 years. Timing shift is 7 years. Relative change of water demand was calculated by formula (12). Using results of calculation graphs of relative change of water demand in time for corn (Fig. 9) were completed. Table 1. Growth of water demand for cropping when irrigating at meteorological station points within the area of the canal (for example, alfalfa and corn) Growth of water demand for the crop, I Si , %

Kobelyaky

Hubinyha

Krasnograd

Izyum

Lozova

Slobo-zhansky

Alfalfa

12.85

3.52

10.91

12.1

11.40

−2

Corn

21.86

9.03

19

19.85

21.9

7.35

98

P. Kovalchuk et al.

Fig. 9. Relative change of water demand for corn using moving average method when interval shift m = 7 years at meteorological station points for 1960–2015 period

Analysis of the graphs shows that over the first two interval of time trend of relative change of crop water demand when irrigation is decreasing for majority of meteorological stations while for the last three intervals it is growing. The maximum of growth is observed within the last interval (1995–2015 period) and it is 18–33%. That means comparing with 1960–1980 period the following two periods were wetter than one and since 1981 drought phenomenon is increasing for the summer (growing season of corn). Evaluation of crop water demand for irrigation in the presence of changing climate in the area of the canal will give the opportunity for agricultural producers to support favorable conditions for irrigated agriculture management, including the prospect of water providing for irrigation.

5 Conclusion The analysis of scientific works made it possible to assess the climatic state and its changes in the region using various criteria and climatic indices. However, the existing approaches do not allow studying the change in water demand for irrigation of agricultural crops with increasing aridity of climatic conditions. If there are stochastic changes in water demand for agricultural crops, it is important to analyze the intensity of drought and identify trends in changes in water demand during the growing season of each year and over a long-term period. To compare the water demand under conditions of stochastic natural moisture in the region, a method and mathematical models for assessing changes have been developed. The method consists in comparing the water demand of agricultural crops based on the calculation of water supply curves at different time intervals. Numerical calculations of ten-day moisture deficits and water demand norms are carried out on the basis of longterm observations at meteorological stations for temperature and air humidity deficit

Crop Irrigation Demand Modelling Method

99

indicators. The method compares the absolute values of the change in water demand in years of the same calculated water supply and the relative change in water demand for two different periods. The mathematical model, based on the use of the statistical moving average method, is aimed at smoothing the long-term series of irrigation norms for agricultural crops by the criterion of the relative change in water demand in certain periods of time. This makes it possible to assess the trends of changes in water demand in the region over the past decades. The developed method and models are used for systematic assessment of changes in water demand of crops in the zone of influence of the Dnipro-Donbas canal. For simulation and numerical calculations, observations of the meteorological parameters of six meteorological stations in 1960–2015, which are located in this zone, were used. Conducted systemic numerical calculations for the Dnipro-Donbas canal region showed an increase in the criterion of water demand in the summer months with irrigation for corn for grain up to 22%, for alfalfa up to 13% in the period 1988–2015 compared to the period 1960–1987. There is an increase in dry conditions in the summer. The trend of increasing aridity in recent years leads to an increase in irrigation rates. Consequently, with the complex use of water resources in the Dnipro-Donbas canal system, it becomes necessary to clarify the balance calculations. It is proposed to use the method and models in the systems of integrated water resources management in river basins, in the zone of influence of canals to predict water demand for irrigation under conditions of climate change. For these purposes, forecasting of meteorological parameters is carried out and will be used in some regions of Ukraine, in particular in the Inhulets river basin. A systematic model of water resources management for the Inhulets River is currently being developed.

References 1. Zhovtonog, O., Filipenko, I., Demenkova, T., Polishchuk, V., Butenko, Y.: Irrigation planning taking into account climate change and droughts intensity in the steppes zone of South Ukraine. Land Reclam. Water Manag. 107, 37–46 (2018). (in Ukrainian) 2. Kulbida, M.I., Elistratova, L.O., Barabash, M.B.: The current state of Ukraine’s climate. Problems of environmental protection and ecological safety. Raider, Kharkiv (2013). (in Ukrainian) 3. Palmer, W.C.: Meteorological drought. U.S. Research Paper No. 45. US Weather Bureau, Washington, DC (1965) 4. McKee, T.B., Doesken, N.J., Kleist, J.: The relationship of drought frequency and duration to time scales. In: Proceedings of the 8th Conference on Applied Climatology, Anaheim, CA, USA, 17–22 January, pp. 179–184 (1993) 5. Yermolenko, N.S., Khokhlov, V.M.: Comparison of spatial and temporal characteristics of droughts in Ukraine at the beginning and at the end of the 20th century. Ukrainian Hydrometeorol. J. 10, 65–72 (2012). (in Ukrainian) 6. Narasimhan, B., Srinivasan, R.: Development and evaluation of soil moisture deficit index (SMDI) and evapotranspiration deficit index (ETDI) for agricultural drought monitoring. Agric. For. Meteorol. 133(1–4), 69–88 (2005) 7. Mika, J., Horvath, S.Z., Makra, L., Dunkel, Z.: The palmer drought severity index (PDSI) as an indicator of soil moisture. Phys. Chem. Earth Parts A/B/C 30(1–3), 223–230 (2005)

100

P. Kovalchuk et al.

8. Selyaninov, G.T.: The origin and dynamics of droughts. In: Drought in the USSR. Their Origin, Repeatability and Impact on Yield, pp. 5–29 Gidrometeoizdat, Leningrad (1958). (in Russian) 9. Zoidze, Y.: On the approach to the study of adverse agro-climatic phenomena in a changing climate in the Russian Federation. Meteorol. Hydrol. 1, 96–105 (2004). (in Russian) 10. Gringof, I.G., Pasechnyuk, A.D.: Agrometeorology and agrometeorological observations. Textbook. Gidrometeoizdat, p. 552 (2005). (in Russian) 11. Tarariko, Yu.O., Saydak, R.V., Soroka, Yu.V., Vitvits’kyy, S.V.: The zoning of the territory of Ukraine by the level of availability of hydrothermal resources and volumes of use of agricultural land reclamation, Kyiv Komprint (2015). (in Ukrainian) 12. Eitzinger J., et al.: Agroclimatic indices and simulation models. In: Agroclimatic Practices and Applications in European Regarding Climate Change Impacts. European Science Foundation; FP 7, ESSEM (2008) 13. Popovich, V.F., Dunayeva, E.A., Kovalenko, P.I.: Using Standardized Rainfall Index (SPI) to assess the level of water availability for the territory and operation conditions of waterworks facilities. Bull. Natl. Univ. Water Manag. Nat. Manag. 2(66), 34–42 (2014). (in Ukrainian) 14. Tadesse, Y., Amsalu, A., Billi, P., Fazzini, M.: A new early warning drought index for Ethiopia. J. Water Clim. Change 9(3), 624–630 (2018) 15. Update of the ICPDR Strategy on Adaptation to Climate Change. International Commission for the Protection of the Danube River (2018) 16. Romashchenko, M.I., Zhovtonog, O.I., Kruchenyuk, V.D., Saydak, R.V., Knysh, V.V.: Managing the restoration and sustainable use of irrigation. Land Reclam. Water Manag. 101, 137–147 (2014). (in Ukrainian) 17. Fischer, G., Tubiello, F.N., Van Velthuizen, H., Wiberg, D.A.: climate change impacts on irrigation water requirements: effects of mitigation, 1990–2080. technol. forecast. soc. chang. 74(7), 1083–1107 (2007) 18. Kovalchuk, P., Kovalenko, R., Kovalchuk, V., Demchuk, O., Balykhina, H.: Integrated water management and environmental rehabilitation of river basins using a system of non-linear criteria. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2020. AISC, vol. 1247, pp. 40–51. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-55506-1_4 19. Arezki, S., Djamila, H., Bouziane, B.: AQUAZONE: a spatial decision support system for aquatic zone management. IJITCS 7(4), 1–13 (2015). https://doi.org/10.5815/ijitcs.2015. 04.01 20. Ahemed, R., Amjad, M.: Automated water management system (WMS). Int. J. Educ. Manag. Eng. 9(3), 27–36 (2019). https://doi.org/10.5815/ijeme.2019.03.03 21. Kovalchuk, P., Rozhko, V., Kovalchuk, V., Balykhina, H., Demchuk, O.: Optimization of integrated water exchange management technologies in territorial systems for conditions of sustainable development. In: 2019 IEEE 14th International Conference on Computer Sciences and Information Technologies (CSIT), Lviv, Ukraine, pp. 80–83 (2019). https://doi.org/10. 1109/STC-CSIT.2019.8929791. 22. Romanenko, V., Zhukynsky, V., Oksiyuk, O.: Methods of environmental assessment of surface water quality according to the appropriate categories. Symvol-T, Kyiv (1988). (in Ukrainian) 23. Mwemezi, K., Sam, A.: Development of innovative secured remote sensor water quality monitoring & management system: case of Pangani water basin. Int. J. Eng. Manuf. (IJEM) 9, 47–63 (2019). https://doi.org/10.5815/ijem.2019.01.05 24. Awad, M., Zaid-Alkelani, M.: Prediction of water demand using artificial neural networks models and statistical model. Int. J. Intell. Syst. Appl. 11(9), 40–55 (2019). https://doi.org/ 10.5815/ijisa.2019.09.05 25. Kovalchuk, P.I., et al.: System modeling and management of water and land use: Monograph. Agrarian Science, Kyiv (2019). (in Ukrainian)

Crop Irrigation Demand Modelling Method

101

26. Allen, R.G., Pereira, L.S., Raes, D., Smith, M.: Crop evapotranspiration – guidelines for computing crop water requirements. FAO Irrigation and Drainage, Paper 56, Rome, Italy (1998) 27. Tsivinsky, G.V., Pendak, N.V., Idayatov, V.A.: Instruction on Operative Calculation of Irrigation Regimes and Forecast of Irrigation of Agricultural Crops due to Lack of Moisture Stores, 2nd edn. Ukrainian Ecological League, Kherson (2010).(in Ukrainian) 28. Romashchenko, M., Bohaienko, V., Matiash, T., Kovalchuk, V., Danylenko, I.: Influence of evapotranspiration assessment on the accuracy of moisture transport modeling under the conditions of sprinkling irrigation in the south of Ukraine. Arch. Agron. Soil Sci. 66(10), 1424–1435 (2019) 29. Climatic Research Unit, University of East Anglia. http://www.cru.uea.ac.uk/. Accessed 1 Sept 2019

The Method of Collision Risk Assessment Using Soft Safety Domains of Unmanned Vehicles Volodymyr Sherstjuk1(B) , Maryna Zharikova1 , Igor Sokol1 , Irina Dorovskaja1 , Ruslan Levkivskiy2 , and Victor Gusev2 1 Kherson National Technical University, Kherson, Ukraine 2 Kherson State Maritime Academy, Kherson, Ukraine

Abstract. This work presents a new method of real-time approximate volumetric collision risk assessment relating to the joint motion of a large group of unmanned vehicles in confined areas. The authors propose the novel concept of multi-level dynamic soft safety domains and spherical dynamic soft topology that allows defining non-spherical safety domains by measuring various radiuses within sectors located in different longitude and latitude. The nonlinearity of the proposed spherical topology allows us to use a novel volumetric approach to collision risk assessments. If the safety domains of various objects overlap at some point within joint motion space, their safety grades should be summed up with respect to the volumes of the cells. The proposed method builds the distribution of the collision risk over the space to prioritize risks properly. The proposed method is quite simple and fast, it allows unmanned vehicles to effectively assess the collision risks imposed by participants of the joint motion process in situations of very dense traffic of numerous vehicles. It provides the acceptable performance of collision risk assessment. The proposed collision risk assessment method is intended to be used in the real-time navigation support systems for large groups of unmanned vehicles to keep safety during cooperative trajectory planning and re-planning. Keywords: Collision risk assessment · Spherical soft topology · Unmanned vehicle · Safety domain · Safety grade · Volumetric approach

1 Introduction Unmanned systems are increasingly being used to solve practical problems in various fields of human activity. Since the cost of modern unmanned vehicles (UVs) is gradually decreasing while their environment is increasing as well as the scope of applications and autonomy grows more and more, a good idea is to use UVs massively in large groups. Autonomous UVs become more intelligent and require less operator’s control making decisions on their own. Thus, they can be effectively united in large teams to solve problems cooperatively. This no longer looks fantastic for today because swarms of drones are successfully used to solve many tasks in the military and other fields [1]. More complex models of cooperative use of UVs have been also proposed, and a good example is the smart fishing operation. During such an operation, a multitude of unmanned aerial © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 102–116, 2021. https://doi.org/10.1007/978-3-030-80472-5_9

The Method of Collision Risk Assessment

103

vehicles simultaneously performs search missions looking for fishing schools from the flight altitude, while plenty of unmanned underwater vehicles move towards the schools of fish found to identify fish species in schools, driving fish into the fishing gear, assess a catch, etc. At the same time, unmanned surface boats carry fishing gear, hook fish and transport the catch to the trawlers. Thus, smart fishery operations involve multitudes of aerial, surface, and underwater vehicles that cooperatively perform different missions in small groups, in pairs, or individually. Such a large group of performers can be organized in the complex structure called an ensemble of unmanned vehicles [2], where each participant plays a certain role in performing a given scenario. During the smart fishing operation, UVs operate synchronously and move jointly in uncertain and dynamic environments, so they need proper navigation support. The most essential challenge is efficient joint path planning [3], which can be solved using cooperative planning of the vehicles’ trajectories that keeps their safety. Since the environment is dynamic, there might be several dynamic (winds, currents, waves, etc.) and situational (obstacles, moving objects, elusion of the fish schools, etc.) disturbances in the process of joint movement along the planned trajectories [4]. Usually, such disturbances and given navigational restrictions (velocity, acceleration, rudder angle, etc.) affect planned trajectories forcing UVs to maneuver. Such maneuvers change the trajectory of the certain UV and provoke other UVs to change their trajectories due to safety reasons. Basically, such operations of large UV groups as smart fishing are carried out in confined areas, so maneuvering can lead to collisions. Thus, the joint motion of a large group of UVs in confined areas is risky enough, where the risk of collisions is caused by a high traffic density. Today, studying the methods of cooperative trajectories’ planning is one of the actively researched areas [5]. This paper is mainly focused on issues of the joint planning of the trajectories of a large group of UVs operating simultaneously within a confined area. In such conditions, the pilot is traditionally responsible for the collision risk assessment using well-proven methods, techniques, and regulations [6]. However, autonomous UVs need to make path planning decisions assessing risk in real-time on their own. Since the UVs’ control systems should cooperatively plan trajectories at risk, there is a serious problem of decision-making because well-known methods of path planning at risk are principally human-centric and cannot help to make such decisions. Thus, the paper’s objective is to develop a method of real-time collision risk assessment relating to the joint motion of a large group of UVs in confined areas.

2 Review of Collision Risk Assessment Methods Risk has been differently defined depending on the purpose but generally, the risk is about some undesired consequences caused by exposure to hazard as well as a chance of being exposed to the hazard. In most cases, risk can be defined as a product of the probability and the magnitude of losses caused by a hazard [7]. We consider potential collisions are hazards. Thus, the risk implies a certain exposure of UV (or a group of UV) to the hazard of a potential collision, which can result in not only such negative consequences as damaging and/or destroying UVs but also contamination of the environment (leaking fuel and oil), losing of fish catch due to mission failure,

104

V. Sherstjuk et al.

and ever losing life or getting severe injuries of people involved in the smart fishing operation. In the context of the paper, the risk and the safety are closely related concepts because the UV’s motion is safe if and only if there is no risk of collisions or the assessed level of risk is below the acceptable minimum. Thus, the collision risk assessment can be grounded on a safety measure (i.e., safety degree). Hazards should be considered as highly unlikely but possible events that are rare but have severe consequences [8]. In this case, the collision risk is always present within a large group of jointly moving UVs. Obviously, such risk is spatially distributed and can occur at any point in time and space. Risk assessment issues have been well studied in the context of manned vehicles, especially much attention has been paid to the methods of the risk assessment associated with human roles in hazardous events and correspondent human errors [9]. It is clear that such methods are weakly applicable to autonomous unmanned vehicles. Traditionally, researchers distinguish four types of collision risk: the real collision risk, the statistical risk of collision occurrence, the predicted risk, and the perceived risk [10]. Usually, UVs cannot assess statistical, predicted, or perceived risks due to a lack of data. Thus, we should address proactive risk assessment models and methods. Such methods can be qualitative or quantitative [11], but in the context of the joint motion of a group of UVs, the quantitative methods are primarily applicable because the UV’s intelligent control system will not be able to provide a qualitative risk assessment in a finite time due to the nonlinearity of the correspondent methods. Therefore, if the human-centric methods that depend on the pilot’s decision will be discarded, there are two suitable groups of the quantitative methods: causal and geometrical. Causal methods for the risk and safety assessment [12] provide definitions of various factors that might cause collisions based on the probability estimation of collision occurrences. This group of methods includes Fault Tree Analysis, Common Cause Analysis, Event Tree Analyses, Scenario Analysis, and other methods that can use Bayesian Belief Networks, Petri Nets, Monte Carlo simulation technique, etc. A good overview of such methods is presented in [13]. The advantage of these methods is the well-grounded prediction of the risk level. Their drawbacks are concerned with the complex modeling process, expert-dependent estimations of safety measures and probability distribution, and essential non-linearity that prevent their use to assess the collision risk for unmanned vehicles in real-time. Geometrical collision risk methods are mainly based on a certain separation of vehicles using certain space and time limits (minima) that protect UVs from conflicts and collisions [14]. Such methods are primarily based on the geometry of vehicles’ approaching. There are several risk assessment models proposed such as the random position/speed deviation model, the intersection model, the geometric conflict model, the model of minimum separation rules, the violation-based model, etc. A good analysis of corresponding methods is presented in [15]. These methods are computationally efficient but most of them are based on a too weak assumption that vehicles follow pre-determined trajectories at constant speeds. In the context of this paper, such an assumption is a serious drawback since the environment is dynamic and the appearance of various trajectory disturbances is quite probable.

The Method of Collision Risk Assessment

105

However, it is advisable to pay much attention to violation-based methods [16]. These methods are based on the well-defined boundaries of specified safety areas (also known as safety domains), which must be tested on the violation. Their main advantages are that they do not require significant amounts of computation; they are quantitative and sufficiently scalable. However, they require well-defined permissible and minimum boundaries of the safety domains, so they completely depend on the availability of objective (at least approximate but expert-independent) definition of the safety domains’ boundaries [17]. Summarizing the review, we conclude that all methods, which rely on statistics or the subjective opinion of experts, are unacceptable in the context of the research at hand since unmanned vehicles have neither statistics nor experts. The methods of Artificial Intelligence are mainly implemented by exhaustive algorithms and are principally nonlinear, so it is impossible to guarantee a finite risk-assessment time. Thus, it is reasonable to develop the violation-based risk assessment method using the geometric approach. Our main assumption is that the UV’s collision risk assessment is not concerned with the precise outcomes because it aims at prioritizing collision risks in the order of their mitigation [18]. Besides, most information captured by UV sensors is numerical but has limited accuracy and can be blurred, therefore, the developed risk assessment method must process incomplete and inaccurate information properly. Finally, to achieve the research objective the computational complexity of the developed method should be relatively low to provide the collision risk assessment in real-time.

3 Methodology We consider a three-dimensional space C of joint motion of a group of UVs. To reduce the computational complexity of the model, we propose to use a soft set [19], which helps us to avoid most of the problems related to both fuzzy and rough sets, which are typically subjective and computationally hard [20]. The soft set enables us to approximately define safety conditions. We propose also define safety domains using topologies, which in our case will be soft. Considering the well-known features of the safety domain model, it is advisable to apply non-traditional topology models to describe them, and in our case, we propose to use a spherical topology [21]. We assume that using such approaches will enable abandoning iterative calculations to reduce the overall computational complexity of the proposed model. 3.1 Topology of Joint Motion Space Suppose the space C is linear and uniform. Let Y be a set of certain elements, and T be a set of time points t strictly ordered by 0 wt,d = (1) 0, otherwise If we use this formula to transform the existing term frequency results, we can see, that in weighted tf the terms frequency was dampened. The weighted tf values are more comparable to each than values for original term frequency. The table below demonstrates the transformation of original term frequencies [12] (Table 2). Table 2. Weighted and original frequency of the terms Term frequency Weighted term frequency 0

0

10

2

1000

4

3.3 Vector Space Model Once we calculate TF-IDF we need to determine which items are closer to each other. This can be done through Vector Space Model. Using VSM we can compute the proximity which relies on the angel between vectors. In VSM each object is stored as a vector in a n-dimensional space. The space between the vectors is calculated to determine the similarity between these vectors.

Fig. 4. The similarity between the vectors

On Fig. 4 we can see the representation of two attributes Security and Education. We have two documents (in our case representation of the content for the website) and two users, who liked one of the topics more that another. To determine and calculate user presences we taking the cosine of the angle between the user and document vectors (User i ).

Content-Based Machine Learning Approach

123

The main reason of using cosine for this calculation is that value of it will increase with decreasing the value of the angle between. This fact signifies more similarity. After normalizing of the vectors, they become the vectors of length 1, that means that calculation of the cosine is simply sum-product of these vectors cos(document1, document2).

4 Methodology We can apply content-based recommendation system model to security mechanisms. The web-based system developed in the frame of our research uses built-in filtering mechanism to calculate term frequency from user input and provides the end user with the relevant content. The system is oriented on analyzing user input. Based on the data entered by the user, system makes calculations using TF-IDF formula to find similarities with the data entered by other users previously. If user has the processor produced by Intel and network card TP-Link, the system will find similarities based on the content and provide the user with recommendations based on particular scenario [13] (Table 3). Table 3. Term frequency based on user input User inputs

Intel CPU TP link

Information block 1

4

5

3

Information block 2

5

2

1

Information block 3

2

3

0

Information block 4

0

1

1

Information block 5

1

2

2

Document frequency 550

1000 400

For example, at the moment user entered data, system keeps information about different processors in 1700 block documents, and Intel is mentioned in 550 documents. In one particular information block 1, word “Intel” appears 4 times and the total number of words in information block 1 is 1 000. Let us assume that total number of information blocks is 22 000. The system will calculate two elements: term frequency and inverse document frequency for each term of user input. Weighted term frequency for “Intel” in information block 1 is calculated as follows: 1 + log10 4 = 1.602

(2)

At the same way term frequency is calculated for other attributes of the information blocks. These values make up the attribute vector for each information block (Table 4). Once the system has term frequency, the inverse document frequency can be calculated by taking the logarithmic inverse of the document frequency amid the whole range of information blocks. So, if there is a total of 22 000 information block documents are

124

G. Iashvili et al. Table 4. Weighted term frequency based on user input User inputs

Intel

CPU

TP link

Information block 1 1.602 1.698 1.477 Information block 2 1.698 1.301 1 Information block 3 1.301 1.477 0 Information block 4 0

1

1

Information block 5 1

1.301 1.301

being found in the database and term “Intel” appears in 550 blocks, the IDF for it will be: log10 (22000/5) = 1.602

(3)

At the same way IDF is calculated for each term in the user input described above (Table 5). Table 5. Inverse document frequency based on user input IDF 1.602 1.342 1.740

For example, for information block 1, length vector (LV) will be calculated as follows:  LV1 = 1.6022 + 1.6982 + 1.4772 = 2, 760 (4) So, to get the normalized vector, each term vector must be divided by the document vector length. Normalized vector for term “Intel” in information block 1 is 1.602/2.760 = 1.021

(5)

Once the normalized vectors for the data are obtained, system can find the similarity between information blocks. The cosine values of the data needs to be calculated. Let us take values for two information blocks (Table 6). Table 6. Calculating cosine of information block values Information block 1 1.602 1.698 1.477 Information block 2 1.698 1.301 1 Information block 3 1.301 1.477 0

cos(I 1, I 2) = 1.602 ∗ 1.698 + 1.698 ∗ 1.301 + 1.477 ∗ 1 = 6.406 cos(I 1, I 3) = 1.602 ∗ 1.301 + 1.698 ∗ 1.477 + 1.477 ∗ 0 = 4.591 cos(I 2, I 3) = 1.698 ∗ 1.301 + 1.301 ∗ 1.477 + 1 ∗ 0 = 4.130

Content-Based Machine Learning Approach

125

5 Experimental Results and Discussions As we can see the data between information blocks 1 and 3 are the most similar to data in information blocks 2 and 3. Based on received data, the system can suggest the relevant recommendations to improve the security of the system including installation of special software or update of existing wireframes and components. Based on provided information, the end-user can improve the security level of existing hardware-based system by following the points from the content of recommendation. For example, if user of the system will know, that concrete model of central processor unit produced by Intel corporation has security problem, that is similar to the results of other users, tested the system on the same hardware, consequently these problems can be solved on local level. One of the ways of solving such security issues is use of patches and security updates provided by the vendors. Content-based recommendation engine for hardware-based systems developed in the frame of our research, is a new method of improvement the security of the users based on particular scenario. The main advantage of such system is an active interaction with concrete user cases. The developed system makes the connection between hard to understand security measures and end-users, that fact has a great impact on overall security of the systems, based on hardware components. It needs to be considered that content-based systems cannot work fully with capturing of inter-dependencies and more complex actions from user-side. Comparing to the other works in this field [1–3], which are described in the literature review section, our approach has much better efficiency.

6 Conclusions Considering the fact, that content-based recommendation systems have some limitations, for the future research we are going to analyze different approaches of machine learning and make a hybrid from the systems to get more user-friendly and accurate platform. Together with the huge potential, different content-based recommendation systems may have some limitations for now. And this is a great field of the research for scientist worldwide. Acknowledgment. The work was conducted in the frame of PHDF-19–519 financed by Shota Rustaveli National Science Foundation of Georgia.

References 1. Pazzani, M.J., Billsus, D.: Content-based recommendation systems. In: Brusilovsky, P., Kobsa, A., Nejdl, W. (eds.) The Adaptive Web. LNCS, vol. 4321, pp. 325–341. Springer, Heidelberg (2007). https://doi.org/10.1007/978-3-540-72079-9_10 2. Lops, P., Jannach, D., Musto, C., Bogers, T., Koolen, M.: Trends in content-based recommendation. User Model. User-Adap. Inter. 29(2), 239–249 (2019). https://doi.org/10.1007/ s11257-019-09231-w 3. Roy, P., Chowdhary, S., Bhatia, R.: A machine learning approach for automation of resume recommendation system. Procedia Comput. Sci. 167, 2318–2327 (2020). https://doi.org/10. 1016/j.procs.2020.03.284

126

G. Iashvili et al.

4. Gorakala, S.K.: Building Recommendation Engines. Packt Publishing, Birmingham (2016) 5. Mohidul Islam, S.M., Debnath, R.: A comparative evaluation of feature extraction and similarity measurement methods for content-based image retrieval. Int. J. Image Graph. Signal Process. (IJIGSP) 12(6), 19–32 (2020). https://doi.org/10.5815/ijigsp.2020.06.03 6. Farhan Sadique, M., Rafizul Haque, S.M.: Content-based image retrieval using color layout descriptor, gray-level co-occurrence matrix and k-nearest neighbors. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 12(3), 19–25 (2020). https://doi.org/10.5815/ijitcs.2020.03.03 7. Al-Jubouri, H.: Integration colour and texture features for content-based image retrieval. Int. J. Mod. Educ. Comput. Sci. (IJMECS) 12(2), 10–18 (2020). https://doi.org/10.5815/ijmecs. 2020.02.02 8. Bauskar, S., Badole, V., Jain, P., Chawla, M.: Natural language processing based hybrid model for detecting fake news using content-based features and social features. Int. J. Inf. Eng. Electron. Bus. (IJIEEB) 11(4), 1–10 (2019). https://doi.org/10.5815/ijieeb.2019.04.01 9. Iashvili, G., Iavich, M., Gagnidze, A., Gnatyuk, S.: Increasing usability of TLS certificate generation process using secure design. In: CEUR Workshop Proceedings, vol. 2698, pp. 35– 41 (2020) 10. Achakulvisut, T., Acuna, D.E., Ruangrong, T., Kording, K.: Science concierge: a fast contentbased recommendation system for scientific publications. PLoS One 11(7), e0158423 (2016) 11. Gnatyuk, S., Barchenko, N., Azarenko, O., Tolbatov, A., Obodiak, V., Tolbatov, V.: Ergonomic support for decision-making management of the chief information security officer. In: CEUR Workshop Proceedings, vol. 2588, pp. 459–471 (2019) 12. Smriti Ayushi, V.R., Prasad, B.: Cross-domain recommendation model based on hybrid approach. Int. J. Mod. Educ. Comput. Sci. (IJMECS) 10(11), 36–42 (2018). https://doi.org/10. 5815/ijmecs.2018.11.05 13. Mammadova, G.A., Aghayev, F.T., Zeynalova, L.A.: Use of social networks for personalization of electronic education. Int. J. Educ. Manag. Eng. (IJEME) 9(2), 25–33 (2019). https:// doi.org/10.5815/ijeme.2019.02.03

Diagnosis of Rail Circuits by Means of Fiber-Optic Cable N. Mgebrishvili1 , M. Iavich2(B) , Gr. Moiseev1 , N. Kvachadze1 , A. Fesenko3 , and S. Dorozhynskyy4 1 Georgian Technical University, Tbilisi, Georgia 2 Caucasus University, Tbilisi, Georgia

[email protected] 3 Taras Shevchenko National University of Kyiv, Kyiv, Ukraine 4 National Aviation University, Kyiv, Ukraine

Abstract. For the safety of train traffic, the most important step is the introduction of a new type of rail circuits – fiber-optic rail circuits. The high sensitivity of the fiber optic cable to external influences (deformation, vibration) is an important property both for detection mechanical damage of rails and wheel sets and positioning the rolling stock. The branches of the fiber-optical cable through mechanical amplifiers perform both the functions of the information perception element - the sensor and the conducting channel of the transmitted information. Using OTDR (Optical Time Domain Reflectometer), based on the analysis of the backscattered light signal, the form of the physical impact that caused the bending is determined. By the time between the emission of the light signal and the receiving of the backscattered signal, position of damage is calculated. The novel method offered in the article gives us the opportunity to detect dropped or damaged wheel-set in the rolling stock and it helps us to detect the station-to-station block occupation of the rolling stock. It also gives us the possibility to ascertain precise location of occupation of the station-to-station block by the rolling stock and to determine a place of the worn dropped or damaged rail. By means of this method, we can identify the spoiled rolling stock and detect overheated boxes in the rolling stock. This method also allows us to control the load on the rolling stock axles. The paper illustrates the schematic diagram and the algorithm of the proposed system developed by the authors. Keywords: Fiber optic cable · Light signal · Rail circuits · Sensor · Signal processing

1 Introduction The implementation of large investment programs in the intensification of international transport significantly expands the role of railway transport. The increase in traffic intensity and the number of trains of railway vehicles requires to solve the problems of improving safety of traffic and reliability of the monitoring system. Modern requirements for traffic safety have sharply revealed the problems of rolling stock monitoring and hazard © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 127–137, 2021. https://doi.org/10.1007/978-3-030-80472-5_11

128

N. Mgebrishvili et al.

prediction [1]. One of the main problems of the safe traffic of trains is the monitoring of the interaction of the rail and the wheel set and timely detection of its undesirable results [2, 3]. From this point of view, the stages are the most dangerous object. The main aims of railway engineering is the development of monitoring system, which will be able to detect of the exact location of rolling stock, identifying worn of rails and wheel sets, identifying damaged sections of rails and damaged wheel sets, monitoring of static and dynamic load on axles. A large number of works have been carried out in this direction and many devices have been created. Despite this, the research in this area is becoming increasingly actual.

2 Literature Review The authors of [1–3] discuss the problems of minoring and evaluation of the rolling stock. It must be mentioned, that during many decades, basic controlling means of the main part of the railway infrastructure - railway sections and rails, for controlling respectively their vacancy and intactness, are electric rail circuits or electric track circuits, whose conductors are rails. By its use in fact there is proceeded only the control of vacancy of the block-sections [4]. It forms information by determination of such simplest fact as is existence or non-existence of the signaling current in the railway track. Because of this it is functionally restricted. Namely, it is fixing occupation of the block-section but cannot give information about exact place of rolling stock. For electric track circuit to detect damage of a rail, the rail should be broken in such a way that there is not any electrical contact between wreckages [5, 6]. Though, even realization of these restricted functions by the rail circuit is technically quite complicated and expensive. At the same time, their reliability is not great and eventually they are characterized by such important shortcomings as are: non-reliability and non-stability due to: • low resistance of the ballast; • necessity of including of the throttle-transformers and other field devices for arrangement of canalization of the reverse • traction current and locomotive signalization; • generation of dangerous influence of the traction current; • decentralized disposition of the devices; • restriction of the informativeness of the locomotive signalization and etc. [7]. In spite of the use of the modern digital and communication the practice shows that in many cases the modern electric rail circuits do not meet completely the conditions corresponding to the working regimes [8]. The emergency situations are especially increasing at fuel and energy crisis, in the conditions of sharp changes of the weather and pollution of the ballast with electro-conductor admixtures. The influence of variation of the ballast resistance on the precision of operation of the rail circuit remains as a problem. Damages of the rail circuits caused by the rail joints are not eradicated. The rail circuits cannot detect a crack (damage) or wear of the rail, even more, a wear or damage of the wheel-set. No less problematic is control of the rolling stock in the conditions of movement. The solution of this problem will give the opportunity to protect the rail and the wheel set axle

Diagnosis of Rail Circuits by Means of Fiber-Optic Cable

129

from influence of the overloaded or unevenly loaded carriages and possible emergency situations caused by them. The solution of this problem is especially important for the railways with complex relief. On the way of improvement of the train traffic safety and information management systems, the most important step is the introduction of a new type of rail circuits - fiberoptical rail circuits [9, 10]. Given the obvious shortcomings of existing electrical rail circuits, it is not surprising that since the early 2000s [11], attempts have been made to improve existing control systems by implementation fiber optic cables. In the result of interactions with different research groups and producers of fiberoptical equipment we have find out that on the Turkish railway, the first attempt was made to introduce optical sensors to improve the reliability of monitoring in railway transport. As it was predicted sensors based on fiber Bragg grating (FBG sensors) showed very high sensitivity, but these experiment revealed also several meaningful conclusion, which made our group to refuse from using FBG sensors as a basis for a modern control system. There were two aspects that made us refuse from FBG technology: 1) technology and 2) reliability. 1. FBG technology implies back scattering from every grating on the way of light. So the intensity of light is dropping with passing every FBG sensors. If we set many FBG sensor one after another most of them will appear useless, as the light is not achieving them. 2. Even in the case when used just several FBG sensors there was very complicated picture to analyze. For working monitoring system there is required enormous quantity of such sensors, which leads to complicated software. All this decreases the reliability of the prospective monitoring system [12]. The authors of [13, 14, 18] offer different ways of analyzing the efficiency of Fiber-Optic Cable. The authors of [15–17] offer different design and evaluation approaches of the sensors, which we use in our work. These results pushed us to start develop better ways of executing monitoring.

3 Description of the Proposed Approach Research of properties of fiber-optical cable revealed, that the optical cable exhibits sensitivity to the external physical impact [13]. The high sensitivity of the fiber optical cable to external influence is an important property, since high sensitivity is required to detect thermal expansion of rails, micro cracks and other mechanical damages. The sensitivity of the optical cable is manifested in a violation of the linearity of the light it conducts. With physical impact, the fiber-optic cable is undergoing deformation and vibration. This process is classified into two groups - micro bends and macro bends. Conducted light is scattered at the bend. Part of the scattered light returns to the light emitter. On this principle is built OTDR (Optical Time Domain Reflectometer) device [14]. In fact, the fiber-optical cable at the bend turns into a sensor [15, 16]. On the basis of analysis of the back-scattered signal, the type of impact that caused the bending is established. The time required from sending the signal to receiving the backscattered signal with high accuracy sets the distance to the bend [17].

130

N. Mgebrishvili et al.

The author’s group proposed the creation of a centralized monitoring system using fiber-optical cable. Such a system has no analogues in the world. The main idea is to bring the fiber optic cable in tight mechanical contact with a rail. High sensitivity of the cable allows controlling the whole rail. From every physical impact on the section of cable there appears bending and we can precisely estimate location of this impact. So there, fiber optical cable becomes sensor in the place of physical impact. The system should also detect out damaged wheel-set in the rolling stock, control the load on the rolling stock axles, detect overheated boxes in the rolling stock, determine a place of the worn of damaged rail etc. To achieve these goals, the idea was expressed that it is necessary to increase the sensitivity of an optical cable to recognize various kinds of physical influences. Easiest way is making a “loop” of fiber optical cable with mechanical amplifier in order to increase the area of contact of the optical cable with the rail at a given point. So that, we can place this sensitive “loops" all the way. They will give more diverse information and by associating certain damage with a certain waveform, we can accurately determine the nature of the damage and its location.

4 Methodology There is presented the structure of the identical for both rails system. On the first and second rails, the branches as sensors are located on opposite of each other, with equal distance. In order to simplify the presented system, in the article there is considered description of one, conditionally first rail.

Fig. 1. Block-scheme of the system

On Fig. 1 is presented a block scheme of the proposed system of the first rail: a singlemode fiber-optical cable (1) is tightly fixed along the rail, the branches of which are called

Diagnosis of Rail Circuits by Means of Fiber-Optic Cable

131

“the loops” represent a fragment of a fiber-optical cable with mechanical amplifiers that are fiber-optical sensors (2–1 ÷ 2−n). The task of the “loop” is to increase the sensitivity of the optical cable to external influences. Sensors are located along the rails at such a distance from each other to exclude the possibility of simultaneous operation of two or more sensors. Also, their distance from the OTDR’s are precisely defined. At the beginning of the rail, there is a OTDR with certain wavelengths (3), λ = 1310 nm or λ = 1550 nm, which is included in the circuit of a single-mode fiber optic cable tightly fixed along the rail. Using an OTDR, an optical pulse is generated, and therefore their shape S is determined. Also, there is carried out the exact calculation of the coordinates of the impulse signal power drop (distance from the OTDR to the point of the impulse power drop). (3) OTD DR λ

Direction of train traffic

Fig. 2. The scheme of triggering the pulses of registers (4–1 ÷ 4−n)

When the wheelset of the rolling stock acts on the corresponding sensor (2–1 ÷ 2–n), OTDR (3) begin to generate signals in the form of pulses Sn. The exit of the OTDR is connected with registers (4–1 ÷ 4–n), each of which corresponds to a determinate sensor (2–1 ÷ 2−n). On the Fig. 2 is presented the scheme of triggering the pulses of registers (4–1 ÷ 4–n): Accordingly, to the distance between the pulses (tn interval) of reflectometer (3), signals in the form of pulses (St1 ÷ Stn) the corresponding registers of transmitter (4–1 ÷ 4−n). Also, each register of (4–1 ÷ 4−n) will work when there are sent S signals on the input of the registers. At the same time, these impulses in the form of ISn signals are passed to the computing block (5). The impulses in the form of IISn signals from the structure located on the second rail are sent to the same computing block where the calculation process occurs according to the following logic: 1. As the result of the passage of the wheel pairs in the computing block (5), in the direction of movement, the number of pulses from the first five registers of the both rails are counted by the counters I4a1 ÷ I4a5 and II4a1 ÷ II 4a5, then there is comparison of the quantity of counted impulses, for instance, I4a1 = I4a2 = I4a3 = I4a4 = I4a5 = II4a1 = II4a2 = II4a3 = II4a4 = II4a1, then the error as presented

132

N. Mgebrishvili et al.

I4a1and II4a1 is eliminated and final result is formatted. By the previously developed algorithm there is carried out: determining the number of rolling stocks; determining the speed of the rolling stock, determining the type and number of carriage in the rolling stock, determining the length of the rolling stock; 2. The comparison of the analogue signals, signal S with the etalon S received from the relevant register of the transmitter: from the OTDRs (3) of first and second rails, and allocation of distorted pulses and their subsequent supply to the analysis block (6); 3. In the analysis block (6), the analysis of the received signals’ [18–20] forms are processed by the following logic: a) On the short cd section, the IS, IIS signals from the first and second rail sensors, are getting small together comparing to their etalon – on this section the rail is damaged; b) On the given section, the IS, IIS signals from the first and second rail sensors, are getting small together comparing to their etalon – on this section was moved the heavy weight rolling stock; c) On the given section, the first and second rail sensors (2–1 ÷ 2–n) are detecting the similar changes of signals comparing to the etalon – the rolling stock has the carriages of different weight; d) On the given section, the first and second rail sensors (2–1 ÷ 2–n) are detecting the different changes of the I Sn and II Sn during the movement of the specific rolling stock, namely, if I Sn ≈ S and II Sn = S then there is damaged the rolling stock of that side where was found the inequality; 4. By the previously formatted algorithm, with high precise there are defined: the load on the rail, the worn rail, the worn rolling stock and wear quality; 5. From the analysis block (6) the information is supplied to the indication block (7).

5 Algorithm of Functioning Systems On the Fig. 3. and Fig. 4 are presented the block-schemes of working algorithm of computing block (5) and analysis block (6). On Fig. 3 is presented working algorithm of computing block (5): 1. 2. 3. 4. 5.

6. 7.

From registers (4–1 ÷ 4−n) of I and II rails, input of signals ISn and IISn; Calculation of received from each register signals; Comparison of quantity and elimination of errors; Generating of wheelsets quantity; From proper register of sensor: Comparison of the ongoing from the I and II rails reflectometers (3) the analogue signals S with the etalon S signal and transmitting different (modulated) impulses to the analysis block (6); Calculation of train movement speed between successive sensors on known distance x, and interval τ at passage of wheelsets between these sensors; Displaying values of speeds;

Diagnosis of Rail Circuits by Means of Fiber-Optic Cable

133

8.

Calculation of quantity of received impulses a including existing berween them time intervals τ by quantity of carriages N = F(a, τ) and determiantion of type; 9. Displaying the quantity and type of the carriages; 10. Summation of lengths of received quantity of carriages and calculation of length of trains; 11. Displaying of trains lengths.

Fig. 3. The block scheme of the algorithm of the computing block (5)

On Fig. 4 is presented the description of operation of block-scheme of algorithm of analysis block (6):

134

N. Mgebrishvili et al.

Fig. 4. Presents a block scheme of the working algorithm of the analysis block (6)

1. 2. 3. 4. 5. 6.

Inputting values ISn and IISn from computing block; Detention of damaged rail; Determination of cd segment and calculation of degree of damage; Determination of weight of train; Calculation of total weight of train; Detention of carriages with different weight;

Diagnosis of Rail Circuits by Means of Fiber-Optic Cable

7. 8, 9. 10, 11. 12.

135

Monitoring of train accordingly of carriages weight; Detention of damaged wheelsets; Calculation of degree of whellset damage; Displaying of results of analysis.

6 Experimental Results and Discussions As we can see the offered scheme is rather efficient, but it needs the corresponding experiments to be evaluated. For now, we do not have the corresponding sensors to evaluate the results. We use different simulation approaches to check these results, as the sensors we use raspberry pi devices. During the test of 1000 wheel pairs, where 160 wheel pairs and 80 parts of the rails circuits were spoiled the simulation model identified 120 spoiled wheel pairs and 72 parts of the spoiled railed circuits. From 84 overweighed wagons, 72 were identified. For now, we are working on the improving of the implementation and design of our simulated laboratory [20–22] (Tables 1 and 2). Table 1. Assessment results of the wheels and wagons Wheel pairs – 1000

Wagons - 250

Spoiled

160

Overweighed

84

Identified

120

Identified

75

Table 2. Assessment results of the railed circuites Railed circuits Spoiled parts

80

Identified

72

7 Conclusion A review of existing patents also proves the obvious advantage of the proposed method of a universal system, characterized by high measurement accuracy, advanced functionality and simplified design compared to existing methods. The novel developed schematic diagram of the multifunction systems based on fiber optic cable, will give us the possibility to: • detect out or damaged wheel-set in the rolling stock; • detect without fail the station-to-station block occupation of rolling stock;

136

N. Mgebrishvili et al.

• ascertain precise location of occupation of the station-to-station block by the rolling stock; • determine a place of the worn out or damaged rail (rail); • identify the rolling stock; • detect overheated boxes in the rolling stock; • control the load on the rolling stock axles. Equipping the railway sections with these systems will significantly increase the economic efficiency of transportation, which in turn leads to an increase in the region’s budget revenues. The developed design scheme can also be successfully used to monitor the status of strategic facilities. The experimental results show that the offered scheme is rather efficient, but needs the improvement. We are going to make the dataset of the correct results using the simulation process. In the future we will use this dataset to train our model user machine learning algorithms, it will greatly improve the correctness of the model. Acknowledgement. This work was supported by Shota Rustaveli National Science Foundation of Georgia (SRNSFG) [grant number FR-18-4002] and CARYS 2019 [CARYS-19-121].

References 1. Bykadorov, S.A.: Current problems of high-speed railway traffic. In: Proceedings of Siberian State University of Railway Engineering, p. 36–48(2018) 2. Karpuschenko, N.I., Kotova, I.A., Likratov, Yu.N., Surovin, P.G., Antereykin, E.S.: Interaction of wheels and rails in curved sections. Way Track Econ. 6, 2–5 (2008) 3. Shakina, A.V., Bilenko, S.V., Fadeev, V.S., Shtanov, O.V.: The study of the mechanisms of wear of rails. Fundam. Res. 4, 1103–1108 (2013) 4. Polevoy, Y.I.: Basics of railway automatics and telemechanics. Samara, pp. 63–66 (2006) 5. Mgebrishvili, N., Tatanashvili, M., Nadiradze, T., Kekelia, K.: Increase of railway transportation safety by a new method of determination of wheel pair and rail wear and damage In: 2007 ASME, RTDF 2007-46024 Technical Conference and Bearing Research Symposium – Chicago, Il, USA, 11–12 September 2007. www.asmeconferences.org/RTDF2007/. 6. Mgebrishvili, N., Garishvili, I., Dundua, A., Kutateladze, K., Kutubidze, N., Mghebrishvili, G.: New method of determination of wheel pair’s and rail’s damage. In: Proceedings of Mechanics The International Scientific Conference, Tbilisi, pp. 187–197 (2016) 7. Babaev, M.M., Grebenyuk, V.Y.: Analysis of the influence of electromagnetic factors on the operation of track circuits. Coll. Sci. Pap. Donetsk Inst. Railway Transp. 28, 75–82 (2011) 8. Kirilenko, A.G., Pelmeneva, N.A.: Electric rail circuits - studies. Allowance, 94 p. Publishing House of Far Eastern State Transport University (2006) 9. Kulchin, Yu.N., Kolchinsky, V.A., Kamenev, O.T., Petrov, Yu.S.: Fiber-optic measuring network based on quasi-distributed amplitude sensors for recording deformation. Quantum Electron. 43(2), 103–106 (2013) 10. Glyuk Martin (DE), Myuller Matias (DE) – Rail measuring system - Patent RU No. 2 672 772 (2006) 11. Fiber Optical Sensors for High-Speed Rail Applications -Final Report for High-Speed Rail IDEA Project 19 (2005). http://www.nationalacademies.org/trb/idea

Diagnosis of Rail Circuits by Means of Fiber-Optic Cable

137

12. Marmatek measurement technologies. https://marmatek.com 13. Integrated Security Systems Design. 2nd edn., pp. 327–334. Thomas Norman, Elsevier Academic Press, (2014) 14. Anderson, D.R., Johnson, L., Bell, F.G.: Troubleshooting Optical-Fiber Networks: Understanding and Using Your Optical Time-Domain Reflectometer. 2nd edn., pp. 59–72. Elsevier Academic Press (2004) 15. Ugale, S.P., Mishra, V.: Optimization of fiber Bragg grating length for maximum reflectivity. In: IEEE International Conference on Communications and Signal Processing (ICCSP), pp. 28–32 (2011) 16. Morozov, O.G., et al.: Synthesis of Two-Frequency Symmetrical Radiation and Its Application in Fiber Optical Structures Monitoring. Fiber Optic Sensors, pp. 137–164. InTech, Rijeka (2012) 17. Bai, Q., et al.: Recent advances in brillouin optical time domain reflectometry. Sensors 19(1862), 45 (2019) 18. Laferrier, J., Lietaert, G., Taws, R., Wolszczak, S.: JDSU-reference guide to fiber optic testing, pp. 18–21 (2007) 19. Gnatyuk, S., Zhmurko, T., Falat, P.: Efficiency increasing method for quantum secure direct communication protocols. In: Proceedings of the 2015 IEEE 8th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications (IDAACS’2015), Warsaw, Poland, 24–26 September 2015, vol. 1, pp. 468–472 (2015) 20. Mubasher Hassan, M., Rather, G.M.: Free space optics (FSO): a promising solution to first and last mile connectivity (FLMC) in the communication networks. Int. J. Wirel. Microwave Technol. (IJWMT) 10(4), 1–15 (2020). https://doi.org/10.5815/ijwmt.2020.04.01 21. Korchenko, O., Vasiliu, Y., Gnatyuk, S.: Modern quantum technologies of information security against cyber-terrorist attacks. Aviation 14(3), 58–69 (2010) 22. Mubasher Hassan, M., Rather, G.M.: Centralized relay selection and optical filtering based system design for reliable free space optical communication over atmospheric turbulence. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 12(1), 27–42 (2020). https://doi.org/10.5815/ijc nis.2020.01.04

Method of Detecting a Fictitious Company on the Machine Learning Base Hrystyna Lipyanina , Svitlana Sachenko , Taras Lendyuk(B) Vasyl Yatskiv , and Oleksandr Osolinskiy

, Vasyl Brych ,

West Ukrainian National University, Lvivska Str., 11, Ternopil 46000, Ukraine [email protected], {v.brych,vy}@wunu.edu.ua

Abstract. The role of fictitious firms, which are also conventionally called “anonymous commercial structures”, is very high in committing economic crimes. The problem of their detection is complicated by the fact that they break the chains of economic and financial relations between real enterprises, banks, insurance companies, exporters and importers, trade enterprises. To solve this problem, the authors have developed a method based on machine learning, which detects fraudulent activities of a fictitious enterprise (sham business). Based on the analysis of the existing relevant publications the parameters of detecting the fictitious enterprises are formed. As a result of modeling the economic activity of 100 Ukrainian enterprises, 20 fictitious enterprises were identified by the methods of Support Vector Classifier, Stochastic Gradient Decent Classifier, Random Forest Classifier, Decision Tree Classifier, Gaussian Naive Bayes, K-Neighbors Classifier, Ada Boost Classifier, Logistic Regression. The results of experimental studies have shown that all selected methods of classification have an acceptable result. However, the best are Support Vector Classifier, Gaussian Naive Bayes, Logistic Regression with a forecast score of 0.98 and a standard deviation of 0.02. Keywords: Fictitious enterprises · Business entities · Classification · Machine learning

1 Introduction Economic crime appeared a long time ago, almost simultaneously with the creation of entrepreneurial activity. However, its influence is especially acute nowadays in a number of countries. in particular in Ukraine, when a huge capital was concentrated in the hands of individual oligarchs, which inevitably leads to an unprecedented rise in corruption and a catastrophic decline in the economy and living standards. At the same time, the role of fictitious firms, which are also called “anonymous commercial structures”, is quite high in committing economic crimes. Fictitious entrepreneurship as an independent crime is at the same time a kind of means of committing a number of other criminal offenses in the economic sphere. The problem of detecting fictitious entrepreneurship is complicated by the fact that the fictitious enterprise breaks the chains of economic and financial relations between © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 138–146, 2021. https://doi.org/10.1007/978-3-030-80472-5_12

Method of Detecting a Fictitious Company

139

real enterprises, banks, insurance companies, exporters and importers, trade enterprises, including in settlements for goods. In addition, the investigation of economic crime by law enforcement officers requires a considerable time. Besides, we observe growing application of computational intelligence and computer modelling in different areas [24, 26, 27]. Therefore, the use of modern information technology to identify a fictitious enterprise based on machine learning (ML) is undoubtedly relevant.

2 Related Work The widespread prevalence of economic crimes and their danger to the economy of states have had a significant impact on the need to expand the scope of criminal responsibility and include it in this dangerous type of crime [1], as well as a significant increase in relevant research. For example, in Reference [2] the scale of crimes, as well as the gravity of crimes are considered, and a typical profile of offenders, as well as the CVs of some offenders are presented. In Reference [3], the interaction of legal and illegal actors is analyzed, namely, “international crime” and “cross-border crime”, which is often called “transnational crime” and outlines the consequences of research and policy. The Reference [4] examines the relationship between banks’ offshore activities and the tax evasion of companies that conduct business through these banks. The Reference [5] confirmed that the institutional features of offshore financial centers increase the opacity that pushes firms to avoid taxes through their subsidiaries. The Reference [6] is analyzing the “chain” of illicit money with reference to “tax havens”, starting with their definition, and identified the existing network between tax evasion, money laundering and reinvestment in legal activities. The Reference [7] considers the structure of public involvement in the illegal circulation of cash flows in terms of three aspects: presence in the shadow economy, involvement in corrupt practices and concealment of the share of income aimed at non-payment of taxes. The Reference [8] proposed the integration of the principles of sustainable development, combined with important existing work on social responsibility and corporate taxation, can help achieve the goals of reducing the occurrence and acceptability of tax evasion. In Reference [9], the political and criminological history of the struggle against monetary control is studied. It is shown that on the basis of the obtained empirical evidence the fight against the financial components of the main crimes can be carried out, starting with the problem of money laundering. In Reference [10], the definition of the concept for “legalization” and “laundering” of profit from crime in Ukraine is analyzed. Detection of “fictitious enterprises” should be automated, because it requires the processing of large data sets. For this purpose, the Reference [3] presents the most implicit and widely used algorithms for each step of pre-processing the large data in predictive data exchange, and References [19, 21] explore the technical and contextual capabilities of machine learning algorithms for their application to similar forecasts. In Reference [11] a structured review of single-class support classifiers is proposed, and in References [12, 20] the basic concepts of the multilevel classification strategy,

140

H. Lipyanina et al.

which is called the decision tree classifier, are considered. In Reference [13], a new solution was proposed to perform the exact k-closest classification of neighbors based on Spark. In Reference [14], a security assessment method was developed to detect the weakness of mobile intelligent terminals, based on the Adaboost algorithm, and in Reference [15] three models of the ensemble were proposed. The latter are based on a comprehensive proportional assessment of alternatives, logistic regression, enhanced regression tree, random forest and frequency ratio. In Reference [16], the approach to optimization of gradient descent is offered and the method of stochastic approximation is discussed. In Reference [17] the traditional approaches to the detection of anomalies are considered, namely, the methods based on ML and SBIDS, and in [18, 22] the methods of deep neural networks and ML are considered. [22] proposed a system for predicting student performance using machine learning algorithms: Decision Tree (C5.0), Naive Bayes, Random Forest, Vector Machine Support, K-Nearest Neighbor and Deep Neural Network. [23] presents a classification system that uses the Multi-Filter function selection technique and the Multilayer Perceptron (MLP) to predict defective software modules based on machine learning. In [24], house price forecasting, decision tree classification, decision tree regression and multiple linear regression were performed and implemented with the help of the Scikit-Learn machine learning tool. In [25] the mechanism of forecasting the most widely used machine learning algorithms is studied, namely linear discriminant analysis, logistic regression, k-nearest neighbors, random forests, artificial neural network, naive bayes, classification and regression trees, support of vector machines, adaptive reinforcement, and an ensemble model for predicting the six-year graduation of college students using data from the first year of the college. In general, it can be noted that despite the importance of the research, the above References do not offer specific information technologies to identify fictitious enterprises. In our opinion, the use of machine learning is promising, which allows tracking the financial activities of fictitious enterprises that is important in preventing economic crimes. Therefore, a goal of this paper is to develop a method for detecting the fictitious enterprise based on machine learning. This article is devoted to this topic, the rest of which is distributed in the following article. Section 2 discusses the analysis of related work. Section 3 presents a method for identifying a fictitious enterprise based on machine learning. Section 4 presents the conclusions of the study.

3 Materials and Methods 3.1 Method For the purpose of selection of fictitious enterprises, it is necessary to solve the problem of binary classification. For this task, consider 8 different methods of machine learning classification: Support Vector Classifier, Stochastic Gradient Decent Classifier, Random Forest Classifier, Decision Tree Classifier, Gaussian Naive Bayes, K-Neighbors Classifier, Ada Boost Classifier, Logistic Regression. The developed method of detecting a fictitious enterprise on the basis of machine learning can be described by the sequence of the following steps.

Method of Detecting a Fictitious Company

141

Step 1. Submit a request to identify a fictitious enterprise. Step 2. Enter the data that will be entered directly by the user into the system: company code (ID Company – generated automatically in the system), legal address (Address), physical address (FAddress1,… FAddressn), NACE (KVED), names of managers (PIPKER1,… PIPKERn – can be several), photo equipment with geolocation (Foto). All these parameters can be supplemented with new ones and generalize the existing ones, the method is easily adapted. Step 3. Based on the data entered by the user to analyze the following parameters and search for appropriate values from sources of information created using the appropriate API and for photo processing image recognition method, which is the purpose of further research. The main parameters are: the presence in the single database of the register of legal entities and individuals (EDR); availability of VAT, SSC and single tax (P) in the database; timely payment of taxes (PO); availability of settlements with co-agents (K); information on the presence of company executives in the state register of declarations (VKK); availability of licenses according to NACE (L); the presence of criminal cases under Art. 205 of the Criminal Code of Ukraine (K205); the presence of mentions of company executives with keywords: criminal case, corruption, offshore accounts, etc. (ZMI); availability of land at the legal or physical address (ZD); availability of registered trademarks and services, database of industrial marks, database of inventions and other databases of the Institute for Industrial Property of Ukraine (TovZ); availability of issued motor third party insurance policies, MTIBU policy check, motor third party database, search by state car number, check of the status of the Green Card policy for cars owned by the company (SP); availability of cars and their owners issued to the company (A); coincidence of registered cars with insurance policies (A&SP); availability in the database of exporters (E); availability in the stock market database (F); the presence of cars and their owners registered with the company wanted (AR); the presence of weapons of the owners of the company wanted (ZR); the presence of cultural values of the owners of the company wanted (KR); availability of construction licenses in the company (LB); availability of real estate in the company (NM); availability of the company’s website (NS); availability of equipment, recognition of equipment by the available photo and determination of compliance of geolocation with the production address (FR); availability of the company’s social networks and affiliated employees (FC). Step 4. Next, the data is compared with each other, namely: FR with Foto, to determine whether the geolocation matches one of the company’s addresses and whether the photo shows the relevant equipment; SP with A to check whether the registered car companies match the insurance policies. Step 5. Transfer data to the storage database. Step 6. Convert data to binary values. Step 7. Add the fit parameter, which shows the value relative to whether the company is fictitious or not. Of course, machine learning algorithms operate on digital values, so we assign the corresponding discrete values 0 or 1. Step 8. To verify the correctness of the data display, you need to display (visualize) the previously defined basic statistical characteristics of each numerical feature. Accordingly, we display: the number of values, the average value, the minimum and maximum in values. The std line shows the standard deviation (which measures how scattered the

142

H. Lipyanina et al. Table 1. Descriptive statistics

Parameter

fit

EDR

P

PO

K

VKK

L

K205

ZMI

ZD

Count

100

100

100

100

100

100

100

100

100

100

Mean

0,80

0,82

0,96

0,90

0,55

0,58

0,55

0,82

0,51

0,47

Std

0,40

0,39

0,20

0,30

0,50

0,50

0,50

0,39

0,50

0,50

Min

0,00

0,00

0,00

0,00

0,00

0,00

0,00

0,00

0,00

0,00

25%

1,00

1,00

1,00

1,00

0,00

0,00

0,00

1,00

0,00

0,00

50%

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

0,00

0,00

75%

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

Max

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

Parameter

E

F

AR

ZR

KR

LB

NM

NS

FF

FC

Count

100

100

100

100

100

100

100

100

100

100

Mean

0,52

0,42

0,87

0,85

0,91

0,49

0,54

0,47

0,48

0,46

Std

0,50

0,50

0,34

0,36

0,29

0,50

0,50

0,50

0,50

0,50

Min

0,00

0,00

0,00

0,00

0,00

0,00

0,00

0,00

0,00

0,00

25%

0,00

0,00

1,00

1,00

1,00

0,00

0,00

0,00

0,00

0,00

50%

1,00

0,00

1,00

1,00

1,00

0,00

1,00

0,00

0,00

0,00

75%

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

Max

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

1,00

values are). 25%, 50% and 75% of the rows show the corresponding percentages of values in the corresponding parameter. With the correct visualization of data, it is clear trends and patterns, the ratio of variables, which makes it possible to notice trends very well. Therefore, graphs of data density are also displayed. Step 9. Before studying data based on machine learning, one of the most important steps should be performed – data distribution. The data are divided into two groups: training set 80%, test set 20%. The training set was used to build machine learning models, and the test set was used to assess the quality of the model performance evaluation on unfamiliar data. Step 10. Evaluating the machine learning model can be quite complex. Usually, the model is estimated based on the error value. However, this method is not very reliable, as the accuracy obtained for one test set may differ significantly from the accuracy of another test set. Step 11. For the selected test set, cross-checking and evaluation were performed based on the following eight methods: Support Vector Classifier, Stochastic Gradient Decent Classifier, Random Forest Classifier, Decision Tree Classifier, Gaussian Naive Bayes, K-Neighbors Classifier, Ada Boost Classifier, Logistic Regression. For a correct comparison of machine learning algorithms, each of them is evaluated on the same data.

Method of Detecting a Fictitious Company

143

3.2 Experimental Results and Discussion Python was selected to perform machine-based data analysis. The following libraries were employed: pandas, numpy, train_test_split, SVC, GridSearchCV, SGDClassifier, RandomForestClassifier, DecisionTreeClassifier, GaussianNB KNeighborsClassifier, AdaBoostClassifier, LogisticRecoreion, KF. As the input data, the 100 companies operating in Ukraine were used (20 of them were identified as fictitious). All data are translated into logical binary values, which is clearly seen from the descriptive statistics (Table 1) and the graph of the density of indicators (Fig. 1).

Fig. 1. Indicator density diagram

The density diagram (see Fig. 1) shows that all indicators are dense and have no gaps, which means that the data do not require pre-processing and cleaning, respectively. Therefore, we will cross-check (Table 2) and evaluate based on 8 different methods: Support Vector Classifier, Stochastic Gradient Decent Classifier, Random Forest Classifier, Decision Tree Classifier, Gaussian Naive Bayes, K-Neighbors Classifier, Ada Boost Classifier, Logistic Regression. Table 2 shows that all methods showed a fairly good result, but the best are Support Vector Classifier, Gaussian Naive Bayes, Logistic Regression, with a forecast score of 0.98 and a standard deviation of 0.02, which is a very good result. This result can be seen in Fig. 2, where it is clear that these three methods are closest to 1 and have no scope of error.

144

H. Lipyanina et al. Table 2. The results of cross-evaluation

№.

Method

Forecast assessment

The standard deviation

1

Support Vector Classifier

0.983333

0.023570

2

Stochastic Gradient Decent C

0.966667

0.023570

3

Random Forest Classifier

0.966667

0.023570

4

Decision Tree Classifier

0.950000

0.040825

5

Gaussian NB

0.983333

0.023570

6

KNeighbors Classifier

0.883333

0.062361

7

AdaBoost Classifier

0.950000

0.040825

8

Logistic Regression

0.983333

0.023570

Fig. 2. Boxplot algorithm comparison

Thus, the developed method differs from analogues [6, 7, 23] in that it makes it possible to create a single software environment for public sector employees to prevent economic crimes, as well as to monitor the financial activities of fictitious enterprises.

4 Conclusions A method for detecting a fictitious enterprise on the basis of machine learning is proposed.The pilot studies were conducted on data from 100 companies operating in Ukraine, 20 of which were identified as fictitious. Eight methods Support Vector Classifier, Stochastic Gradient Decent Classifier, Random Forest Classifier, Decision Tree Classifier, Gaussian Naive Bayes, K-Neighbors Classifier, Ada Boost Classifier, Logistic Regression were studied for classifying the fictitious enterprises.

Method of Detecting a Fictitious Company

145

As the result of experimental studies, the best methods of classification were selected: Support Vector Classifier, Gaussian Naive Bayes, Logistic Regression with a forecast score of 0.98 and a standard deviation of 0.02. In the future, it is expected to conduct a further research by analyzing the detection of fictitious enterprises based on methods: Support Vector Classifier, Gaussian Naive Bayes, Logistic Regression. The last will be employed to develop an algorithm for recognizing images of enterprise equipment with geolocation data processing and creating the appropriate software.

References 1. Ashoura, A.J.: Criminal liability against perpetrators of economic crimes. Int. J. Innov. Creat. Change 9(2), 285–299 (2019) 2. Gottschalk, P.: Economical motive. In: Gottschalk, P. (ed.) The Convenience of White-Collar Crime in Business, pp. 17–34. Springer, Cham (2020). https://doi.org/10.1007/978-3-03037990-2_2 3. Passas, N.: Cross-border crime and the interface between legal and illegal actors. Secur. J. 16(1), 19–37 (2003). https://doi.org/10.1057/palgrave.sj.8340123 4. Chernykh, L., Mityakov, S.: Offshore schemes and tax evasion: the role of banks. J. Financ. Econ. 126(3), 516–542 (2017). https://doi.org/10.1016/j.jfineco.2017.09.004 5. Ben Amar, W., He, L., Li, T., Magnan, M.: The corrosive effect of offshore financial centers on multinational firms’ disclosure strategy. Eur. Acc. Rev. 1–30 (2018). https://doi.org/10. 1080/09638180.2018.1497520 6. Sabatino, M.: Crime treasure Islands: tax havens, tax evasion and money laundering. J. Econ. Bus. 3(1) 1–12 (2020). SSRN: https://ssrn.com/abstract=3530218 7. Osipov, G.V., Glotov, V.I., Karepova, S.G.: Population in the shadow market: petty corruption and unpaid taxes. Entrep. Sustain. Issues 6(2), 692–710 (2018). https://doi.org/10.9770/jesi. 2018.6.2(16) 8. Bird, R., Davis-Nozemack, K.: Tax avoidance as a sustainability problem. J. Bus. Ethics 151(4), 1009–1025 (2016). https://doi.org/10.1007/s10551-016-3162-2 9. Levi, M.: Evaluating the control of money laundering and its underlying offences: the search for meaningful data. Asian J. Criminol. 15(4), 301–320 (2020). https://doi.org/10.1007/s11 417-020-09319-y 10. Shepeluk, V.: Fictitious Entrepreneurship in Ukraine as a way of tax evasion. Sci. Notes KROK Uni. 1(57), 58–65 (2020). (in Ukrainian). https://doi.org/10.31732/2663-2209-202057-58-65 11. Alam, S., Sonbhadra, S. K., Agarwal, S., Nagabhushan, P.: One-class support vector classifiers: A survey. Knowl.-Based Syst. Article 105754 (2020). https://doi.org/10.1016/j.knosys. 2020.105754 12. Swain, P.H., Hauska, H.: The decision tree classifier: design and potential. IEEE Trans. Geosci. Electron. 15(3), 142–147 (1977). https://doi.org/10.1109/tge.1977.6498972 13. Maillo, J., Ramírez, S., Triguero, I., Herrera, F.: kNN-IS: an iterative spark-based design of the k-nearest neighbors classifier for big data. Knowl.-Based Syst. 117, 3–15 (2017). https:// doi.org/10.1016/j.knosys.2016.06.012 14. Wang, F., Jiang, D., Wen, H., Song, H.: Adaboost-based security level classification of mobile intelligent terminals. J. Supercomput. 75(11), 7460–7478 (2019). https://doi.org/10.1007/s11 227-019-02954-y

146

H. Lipyanina et al.

15. Arabameri, A., Yamani, M., Pradhan, B., Melesse, A., Shirani, K., Bui, D.T.: Novel ensembles of COPRAS multi-criteria decision-making with logistic regression, boosted regression tree, and random forest for spatial prediction of gully erosion susceptibility. Sci. Total Environ. 688, 903–916 (2019). https://doi.org/10.1016/j.scitotenv.2019.06.205 16. Theodoridis, S.: Chapter 5 – Online learning: the stochastic gradient descent family of algorithms. In: Machine Learning, 2nd edn. A Bayesian and Optimization Perspective, pp. 179–251. Elsevier (2020). https://doi.org/10.1016/B978-0-12-818803-3.00014-3 17. Hamolia, V., Melnyk, V., Zhezhnych, P., Shilinh, A.: Intrusion detection in computer networks using latent space representation and machine learning. Int. J. Comput.19(3) 442–448 (2020). https://www.computingonline.net/computing/article/view/1893 18. Turchenko, V., Chalmers, E., Luczak, A.: A deep convolutional auto-encoder with pooling – unpooling layers in caffe. Int. J. Comput. 18(1), 8–31 (2019). https://www.computingonl ine.net/computing/article/view/1270 19. Canhoto, A. I.: Leveraging machine learning in the global fight against money laundering and terrorism financing: an affordances perspective. J. Bus. Res. (2020). https://doi.org/10.1016/ j.jbusres.2020.10.012 20. Lipyanina, H., Sachenko, A., Lendyuk, T., Nadvynychny, S., Grodskyi, S.: Decision tree based targeting model of customer interaction with business page. In: Proceedings of the third International Workshop on Computer Modeling and Intelligent Systems (CMIS-2020), CEUR Workshop Proceedings, no. 2608, pp. 1001–1012 (2020). http://ceur-ws.org/Vol-2608/ paper75.pdf. 21. Lipyanina, H., Maksymovych, V., Sachenko, A., Lendyuk, T., Fomenko, A., Kit, I.: Assessing the investment risk of virtual it company based on machine learning. In: Babichev, S., Peleshko, D., Vynokurova, O. (eds.) DSMP 2020. CCIS, vol. 1158, pp. 167–187. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-61656-4_11 22. Vijayalakshmi, V., Venkatachalapathy, K.: Comparison of predicting student’s performance using machine learning algorithms. Int. J. Intell. Syst. Appl. (IJISA) 11(12), 34–45 (2019). https://doi.org/10.5815/ijisa.2019.12.04 23. Iqbal, A., Aftab, S.: A classification framework for software defect prediction using multifilter feature selection technique and MLP. Int. J. Mod. Educ. Comput. Sci. (IJMECS). 12(1), 18–25 (2020). https://doi.org/10.5815/ijmecs.2020.01.03 24. Thamarai, M., Malarvizhi, S.P.: House price prediction modeling using machine learning. Int. J. Inf. Eng. Electron. Bus. (IJIEEB). 12(2), 15–20 (2020). https://doi.org/10.5815/ijieeb. 2020.02.03 25. Kang, Z.: Using machine learning algorithms to predict first-generation college students’ sixyear graduation: a case study. Int. J. Inf. Technol. Comput. Sci. (IJITCS). 11(9), 1–8 (2019). https://doi.org/10.5815/ijitcs.2019.09.01 26. Darmorost, I., Dyvak, M., Porplytsya, N., Martsenyuk, Y., Brych, V.: Convergence estimation of a structure identification method for discrete interval models of atmospheric pollution by nitrogen dioxide. In: Proceedings of the 2019 9th International Conference on Advanced Computer Information Technologies, ACIT 2019, pp. 117–120 (2019) 27. Chereshnyuk, O., Panasyuk, V., Sachenko, S., Banasik, A., Golyash, I.: Fuzzy-multiple approach in choosing the optimal term for implementing the innovative project. In: Proceedings of the 2017 IEEE 9th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, IDAACS 2017, pp. 533–536 (2017)

Method for the Criticality Level Assessment for Crisis Situations with Parameters Fuzzification Andrii Gizun1(B) , Zhadyra Avkurova2 , Vladyslav Hriha1 , Anna Monashnenko3 , Nurbol Akatayev4 , and Marek Aleksander5 1 National Aviation University, Kyiv, Ukraine 2 L.N. Gumilyov Eurasian National University, Nur-Sultan, Kazakhstan 3 National University of Life and Environmental Sciences of Ukraine, Kyiv, Ukraine 4 Satbayev University, Almaty, Kazakhstan 5 Pa´nstwowa Wy˙zsza Szkoła Zawodowa w Nowym S˛aczu, Nowy S˛acz, Poland

Abstract. The impact of crisis situations on the security of state information resources, various institutions, enterprises, organizations and the state as a whole is quite significant. Thus, crisis situations can not only slow down the development of the system that has its influence, but also destroy it in general. It is necessary to take adequate measures of threat and the use of safeguards, which determines the importance of assessing the criticality of the current situation to prevent such an impact. At present, there are no generally accepted universal criteria and an integrated criticality assessment indicator. Therefore, determining the level of criticality of an incident that can cause the crisis situation is an urgent and important scientific task. The study introduces a set of parameters for assessing the level of criticality of the situation, proposed a method for determining the level of criticality using expert approaches and fuzzy logic methods that do not require collection and processing of statistics, and described the procedures for fuzzification defuzzification. The method can help to better assess the criticality of information systems crises and improve protection against attacks, because the result shows the weaknesses of the systems. Keywords: Crisis situation · Incident · Criticality level · Indicator · Criteria · Criticality level assessment · Losses · Criticality class · Expert methods · Fuzzy set theory

1 Introduction Protection of information resources (IR) from the impact of crisis situations CS and their consequences is currently the most urgent task in the field of information security. Any information security incidents have their causes, i.e. destabilizing factors that cause them and always create a negative impact on the management of information resources of the organization or IR. Thus, numerous incidents in the absence of control over their course and appropriate response can have critical consequences. According to the definition of CS given in [1], it is characterized by large losses, serious interruptions of © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 147–161, 2021. https://doi.org/10.1007/978-3-030-80472-5_13

148

A. Gizun et al.

business processes, questioning the possibility of further functioning of the organization, the destruction of the structure of an individual enterprise or the whole industry, potential threats to people’s life and health. Thus, the CS can not only violate the security characteristics of IR (confidentiality, integrity and availability), but also disrupt their management processes, lead to their loss. At the same time, the higher the level of criticality of the CS, the more severe the consequences it can have and, of course, the more effective anti-crisis measures and measures should be. Therefore, in order to take effective countermeasures, the maximum elimination of the consequences, it is necessary to determine the level of criticality of the CS generated by the incident-potential crisis situation (IPCS), taking into account the dynamics of its development. The processes of protection of information resources by the influence of the CS are regulated by the concept of business continuity management (CBCM). It involves monitoring the current situation, forecasting the CS, assessing the level of criticality of the situation, taking countermeasures and eliminating their consequences, and generally corresponds to the stages of the cycle of Schuhart-Deming or PDCA. Each of these processes has its own characteristics and a different degree of implementation in practice. The article consists of 5 sections: Introduction, Literature review, Methodical Foundations of the Method for the Level of Criticality Assessing of Crisis Situations with Parameters Fuzzification, Experimental Study of the Method for the Level of Criticality Assessing of Crisis Situations with Parameters Fuzzification, Conclusions and Future Researches.

2 Literature Review At the moment, the definition of the concept of CS, their classification were considered in works [1, 2], described and developed methods and systems for forecasting, identification of anomalous state in the IR [3, 4], the activities of violators [5, 6], computer attacks [7]. The work [8] is devoted to the issues of identification, forecasting and modeling of emergencies of ecological, social and state security, which highlighted the criteria for modeling CS industrial, natural and social nature, which, however, are not universal and can be applied to the entire set of possible disasters. And in [9] the main indicators of national security are introduced, among them: depopulation rate, level of shadowing of the economy, level of defense, science and education expenditures, crime rate, decile coefficient, but these criteria characterize the condition of state protection. In addition, there are a number of regulations governing the processes of risk analysis, which include determining the probability of occurrence of CS, probable economic damage, human, individual and collective risk. Among the considered methods there are different classes of methods of risk analysis, namely: deterministic, probabilisticstatistical (statistical, theoretical-probabilistic, probabilistic-heuristic), in conditions of uncertainty of non-statistical nature (fuzzy and neural network), combined. However, none of these methods can be used to process the relevant criteria for the criticality of the situation of various kinds, is not universal and does not take into account all the features of any CS. It should also be noted that the features of the management of security information in the uncertainty of destabilizing factors, methods for assessing the performance of security functions, risk assessment and decision-making in the CS, as well as

Method for the Criticality Level Assessment

149

models for counteracting threats to information security with the option depending on the probability of attack and the method of assessing the level of information security based on fuzzy logic [10]. The last two works consider information of security from the point of view of security, and this research - on the contrary from the point of view of criticality of violation of information security. Having considered the advantages and disadvantages of the ideas of scientists, we have developed a new method for assessing the criticality of crisis situations, based on fuzzy logic.

3 Methodical Foundations of the Method for the Level of Criticality Assessing of Crisis Situations with Parameters Fuzzification 3.1 Formulation of the Problem Subsequent paragraphs, however, are indented. The problem of assessing the level of criticality of the CS, as one of the processes of CBCM, is determined by the fact that its emergence and development is difficult to predict (and often not predictable), i.e. we are dealing with an event in a vaguely formalized space. In addition, there are no generally accepted criteria for assessing the level of criticality, most of them have different nature (including clear and fuzzy) and mathematical properties, which makes it impossible to use most of the currently known assessment methods to the general set of these criteria. Therefore, the formation and development of parameters for assessing the level of criticality of the CS and methods for its determination is an urgent task. Therefore, the purpose of this article is to determine the set of parameters for assessing the level of criticality of the CS, to develop methods for assessing the proposed parameters and to calculate the overall level of criticality of the situation. Let’s consider the CS and its impact on the system, organization or state. Thus, the CS can cause changes in the structure, functional processes that threaten their existence and is characterized by the level of criticality (The level of criticality of the situation) LCS, with increasing level of criticality of the incident increases the likelihood of its transition to a state of CS and significant negative impact on systems. You can assess the impact of the CS using the parameters to assess the level of criticality of the current situation. The parameters can be of different nature, characterize the impact of the CS from different sides, so there are problems in their application by known methods of risk analysis, identification of possible consequences and losses. These parameters can be represented qualitatively (as a linguistic change (LC) with a certain number of terms) or quantitatively. We describe a possible set of parameters for assessing the level of criticality, based on the point of view of the maximum universalization of these characteristics and using the provisions on the classification of CS, set out in [2]. It should be noted that when working with specific types of CS should be able to replenish this set with additional parameters. Let’s form a set of parameters, which will be defined below in step 1 of the method of assessing the criticality of the CS. Specific parameters were selected based on the analysis of the main standards of KUBB (BS ISO/IEC 17799: 2005, BS 25999, NIST ST800-34, NFPA 1600 and others), best practices and practices such as DRII, Gartner, BSI, HP, modern management systems of the CS and the above-mentioned works.

150

A. Gizun et al.

The proposed parameters are unclear, as the expert’s assessment is characterized by the function of belonging (FB) to a certain term of fuzzy number (FN) (for example, for the parameter the degree of dysfunction of critical resources/processes - complete, insignificant, significant, etc.) according to his subjective decision, rather than objective reasons, there are no critical values of these parameters, universal for them measurement scales and reference values and the expert’s assessment does not give an unambiguous answer about the criticality of IPCS. That is why when processing these parameters it is necessary (and possible) to use methods of expert evaluation and fuzzy logic. 3.2 Justification of the Method To assess the level of criticality of the CS, we use a fuzzy model with a linguistic scale (FMLS) [11], when based on expert data, reference values are built, and as a result of measuring the current level of each parameter, a decision is made on the overall criticality level of IPCS. To formalize the processes of forecasting, detection, identification and evaluation , where n

of CS, we introduce a set of IPCS:

determines the number of potential CSs and, incidents that can cause a crisis, each of , Tei , P , which is displayed as a generalized six-component tuple [5]: ∼ i

ERi , LCSi >, which: IKSi – the identifier of the another IPCS, which is (or may be) the cause of the CS; Pi – a subsset of possible parameters used to predict or identify the another incident; Tei – a subset of all possible fuzzy (linguistic) standards that reflect the – a subset of the reference states of the corresponding parameters from the subset Pi ; P ∼ i

current values of the parameters for a certain period of time; ERi – a subset of heuristic rules (similar to [7]), built on the basis of fuzzy parameters used to detect/identify the another IPCS; LCSi – the level of criticality of the situation caused by the another IPCS. A detailed description of the procedure for detection, identification of IPCS, described in [11]. The identified situation refers to the crisis only if the level of its criticality is above average or higher, i.e. LCSi ≥ BC e . Otherwise, the incident is either ignored (with a ∼

sufficiently low level of criticality) or responded to in order to control and eliminate as for a normal information security incident. Therefore, after the detection of IPCS, it must be evaluated. We describe a method of assessing the criticality of the situation, which is a consequence of the impact of a particular incident. It also consists of 6 stages, a schematic representation of which is shown in Fig. 1. The method consists of 6 stages: determination of criticality assessment parameters, formation of evaluation standards, calculation of CI, measurement and fuzzification of parameters, calculation of criticality level of CS, visualization of results. Stage 1. Determination of Parameters for Assessing the Level of Criticality. The level of criticality can be described taking into account the functional relationships between Le – parameters for assessing the level of criticality. The parameters Le may have a different nature, characterize the impact of the CS from different sides, so there are problems in their application by known methods of risk analysis, identification of

Method for the Criticality Level Assessment

151

Fig. 1. Method of assessing the criticality of the situation

possible consequences and losses. These parameters can be represented qualitatively (as a linguistic change (LC) with a certain number of terms) or quantitatively. After analyzing modern methods of risk assessment, best practices of KUBB, the E  following set was formed: L = { Le } = {L1 , ..., LE }, for example, due to the condie=1 15 

tions of research at E = 15: L = {

e=1

Le } ={L1 , , ..., L15 }, where L1 = TR, L2 = DVF,

L3 = GS, L4 = OS, L5 = OLED, L6 = RD, L7 = RTLH , L8 = RM , L9 = F,

152

A. Gizun et al.

L10 = DDI , L11 = CRT , L12 = CRP, L13 = LM , L14 = DIEPF, L15 = DVChS identify such parameters as: “Duration of the incident”, “Degree of disruption of critical resources/processes”, “Geographical scale of the incident”, “Scale of the incident in the organizational aspect”, “Total level of economic losses”, “Ratio of economic losses for the current period to appropriate level for the previous period”, “Level of threat to human life and health” , “Specific mortality rate at the moment”, “Frequency of incidents (intensity)”, “Degree of infrastructure destruction”, “Ratio of estimated recovery time and RTO”, “The ratio of the level of resource loss and RPO”, “The level of panic, protest and anti-government sentiment of staff/population”, “The degree of influence of external destabilizing and psychological factors”, “The degree of violation of safety characteristics of ODR with OD”, described in [12]. Stage 2. Formation of Evaluation Standards. During the second stage, evaluation standards are formed, which will be used for comparison with the FN formed during the determination of the level of all parameters and the overall level of criticality (fuzzification). A separate standard is formed for each parameter, but it is possible to use one evaluation standard, which we will do in the future to simplify the presentation of the method. TeEL

={

r  s=1

=

r 

T∼

e ELs

} = {T∼

e EL1

, ..., T∼

e ELr

rs r   e }= { μeELsq /xELsq } s=1 q=1

e e e e {μeijs1 /xijs1 , μeijs2 /xijs2 , ..., μeijsrs −1 /xijsr , μeijsrs /xijsr }, (q = 1, rs ), s −1 s

(1)

s=1

where rs (s = 1, r) – the number of components in T∼ e

ELs

and T ∼

with similar terms as in T ∼

Le s

[13]. We construct this standard using the method of constructing parametric LCSs

FN described in [11]. The function that sets the value of the FN evaluation standards will look like: ⎧ 0, if x < a; ⎪ ⎪ ⎨ (x − a)/(b − a), if a ≤ x < b; μA (x) = . ∼ ⎪ (c − x)/(c − b), if b ≤ x < c; ⎪ ⎩ 0, if x > c. Range of carrier change the fuzzy number with r = 5 terms and rs = 5 the component is represented on the universal set U = [0, 1]. The obtained FN are presented in Fig. 2, and their mathematical description of the expression:

Method for the Criticality Level Assessment

153

Fig. 2. Graphical representation of standard FN

Stage 3. Calculation of the Coefficients of Importance (CI). The stage used for the calculation (CI) and in accordance with the ranking of criticality assessment parameters. We use the method of quantitative pairwise comparison with the definition of the square radical, which is a kind of method of quantitative pairwise comparison [12]. It is based on the formation of a pairwise comparison matrix A = aee , where aij is selected based on the judgments of the expert according to the scale of relative importance: 1 – alternatives have equal value (priority, importance), 3 – experience and judgment give a slight advantage of one alternative over another, 5 – experience and judgment give a strong advantage of one alternative over another (there is convincing evidence in favor of one of the alternatives), 7 – one of the alternatives far outweighs the other, which is obvious, 9 – the advantage of one alternative over another is undeniable and absolute; 2, 4, 6, 8 - compromise cases. If when comparing the first alternative with the second obtained the above number (for example, 5), then when comparing the second alternative with the  first - the inverse value (1/5). Weights are calculated according to the expression E aee , e = 1, E , where E – number of valuation parameters. After that the ωe = n e =1

rationing of the received coefficients on expression is carried out e = ωe /( that

E

E

ωe ) so

e=1

e = 1.

e=1

Stage 4. Measurement and Fuzzification of Parameters. At this stage, the calculation of the FN is made, representing the current values of the parameters measured by the system and fuzzed. The system evaluates the parameters Le according to the reference values. Based on T the current measurements, made over a period of time, FN is formed that reflects the current value of the parameter. It is defined as =( L ∼ e

r  s=1

T∼ E )/T = T∼ E ELs

EL1

∼

+ T∼ E

EL2

∼ ∼ ∼ ∼ + · · · + T∼ E + · · · + .. /T ELs

(2)

154

A. Gizun et al.

where T – total number of measurements, T∼ E – correction standard. T∼ E is determined ELs

ELs

by sensors used to monitor the current state of the parameters, and the meter mechanism. In essence, the procedure is similar to the method of fuzzification of parameters described in [12]. To do this, enter the set of all sensors S and a subset of sensors defined for each given range of values of all the presented parameters Se ⊆ S: Se = {

re 

Seq (tT )} = {Se1 (tT ), . . . , Sere (tT )},

(3)

q=1

where Seq (tT ), (e = 1, E,q = 1, re ) is a sensor Neq of the interval, which displays the value of the parameter Le (tT ) at the appropriate interval at the time tT , a re – the number of sensors equal to the number of intervals. Sensor Seq (tT ) implemented as a binary function that is equivalent to one only if the value Le (tT ) at the moment of time tT will be in the interval Neq that is:  1, if Le (tT ) ∈ Neq · (q = 1, re ), (4) Seq (tT ) = 0, if Le (tT ) ∈ / Neq To determine the frequency of occurrence of values Le (tT ) at each of the intervals Neq , (q = 1, re ) enter the set of all sensor counters CS and a subset of such meters CSe ⊆ CS for each parameter, which determine the required frequencies of the expression: CSe = {

re 

CSeq } = {

re T max 

Seq (tT )},

(5)

q=1 T =1

q=1

where CSeq – counter sensor Seq (tT ), a Tmax corresponds to the total number of possible tT , i.e. the number of measurements [13]. The obtained frequencies are displayed by counters. The next step in the fuzzification procedure [14] is the construction of correction standards based on the readings of the sensor counter and parameters of FN. In order to implement this step, a set of all possible correction standards is introduced TE and a E subset of the corresponding parameters of the correction standards TE e ⊆ T , which are E based on Te and are defined as TE e

={

re 

T∼ E } = {T∼ E , . . . , T∼ E } = {T∼ e · CSe1 , . . . , T∼ e CSere }, eq

s=1

e1

ere

e1

(6)

ere

where T∼ E (s = 1, re ) – correction standards FN with re terms. They are formed by the es

expression: {

re  s=1

at (s = q = 1, re ) [13].

T∼ E } = ..{T∼ e · CSe1 , . . . , T∼ e · CSere }, es

e1

ere

(7)

Method for the Criticality Level Assessment

155

The obtained data make it possible to calculate fuzzy values of parameters over time. So the result of the fuzzification procedure is formed by the expression [13]: =( L ∼ e

re 

∼

∼

∼

T∼ E )/Tmax = (T∼ E + T∼ E + . . . + T∼ E )/Tmax ,

s=1

es

e1

e2

(8)

ere

where ={ L ∼ e

re 

μeq /xeq } = {μe1 /xe1 , . . . , μesre /xere }, (q = 1, re ),

q=1

. re – the number of components in the current FN L ∼ e

Stage 5. Calculation of the Level of Criticality of the CS. At the fourth stage, the calculations of the general assessment of the level of criticality of the situation are carried out. Initially, taking into account the defined CI, the FN is formed E 

S = LC ∼ i

(e ∗ L ) ∼

(9)

e

e=1

The generated FN is compared with the evaluation standard by one of the known methods of comparing the FN. For these purposes, we use the method of forming the α- level nominalization of the FN and the method of determining the identifying terms. The procedure consists in the calculation of nominalized (converted) standards and the level of criticality r  ep (previously broken down into α-level ALELg i ALLCSg ) TEL = { T∼ ep } = {T∼ ep , …, T∼ ep }, where T∼ ep = { ELr

ELs

ep

z  g=1

ep

s=1 ELs ep ep ep ep ={μELs1 /xELs1 ,, ...,μELsz /xELsz },

ep

μELsg /xELsg }

ep

p

(s = 1, r), a μELsg = μELs(z - g+1) = ALELg i xELg = xELq +

EL1

(g = 1, z),

p

(μELg −μELq )(xELq+1 −xELq ) , μELq+1 −μELq

(g = 2, z), z – a number of component T∼ ep [13]. ELs

A similar procedure is performed with the current values of the criticality level. Next, we determine the generalized distance of Heming. h(T∼ ep , L C Sp ) = ∼ ELs

z       ep  p ep p xELsg − xLCSg  = xELs1 − xLCS1  g=1

   ep  ep    ep  p p p + xELs2 − xLCS2  + . . . + xELsg − xLCSg  + . . . + xELsz − xLCSz ,

(10)

where (g = 1,z), (s = 1, r) [13]. The criterion of conformity LC S one of the terms of ∼ the estimation standard is the shortest distance of Heming. Thus, the level of criticality of the situation or IPCS corresponds to the corresponding term. r

h mins = ∧ h(T∼ ep , LC SP) ∼ s=1

ELs

(11)

156

A. Gizun et al.

Stage 6. Visualization of Results. The results obtained in fuzzy form are shown in Fig. 1. In addition, to better display the criticality level of IPCS, it is proposed to display the criticality parameters using the criticality indicator. The appropriate parameters for should be pre-defuzzificated. The most appropriate in this case is the use of the this L ∼ e

center of gravity method [14], in which the FN is converted into a clear formula L = 100 ∗ (

q  i=1

xLq ∗ μ(xLq )/

q 

μ(xLq ))

(12)

i=1

where q – number of FN calipers. It is possible that the values of individual parameters are calculated directly without the use of expert methods. In this case, they are displayed on the indicator by a histogram [13].

4 Experimental Study of the Method for the Level of Criticality Assessing of Crisis Situations with Parameters Fuzzification 4.1 Experimental Technique We have considered the work of the method on a specific example according to the conditions of the study, estimating the overall level of criticality of the situation on the basis of previously entered parameters. The developed method was used to evaluate the criticality of mail server failures due to a DDOS attack, which took place at the National Aviation University (Kyiv, Ukraine) in October 2020. As the parameters L6 and L8 are clear, then at this stage they remain unchanged. As a result of ranking the parameters, a set of CS values is obtained, which are shown in Table 1. The next step is to measure the value of the controlled parameters and the implementation of fuzzification. The essence of fuzzification is to process the current values of the controlled parameters, taken over a period of time, and presented in the form of a single FN. The fuzzification procedure is performed on the basis of the method described in [13, 14]. Only fuzzy parameters are fuzzed, namely at the stage of assessing the level of criticality – L1 = TR, L2 = DVF, L3 = GS, L4 = OS, L5 = OLED, L7 = RTLH , L9 = F, L1 0 = DDI , L1 1 = CRT , L1 2 = CRP, L1 3 = LM , L1 4 = DIEPF, L1 5 = DVChS. Let’s consider the above-described procedure of fuzzification on the example of the parameter “The degree of influence of external destabilizing and psychological factors”, L14 = DIEPF. When forming the standard parameter was introduced 3  N14q } = {N141 , N142 , a set of identifiers of the intervals of its values N14 = { q=1

N143 } = {NDIEPF1 , NDIEPF2 , NDIEPF3 } = {[0; 20], [20; 50], [50; 100]} and the following reference terms for the same evaluation parameter = T∼ e H ∼

DIEPF1

C = T∼ e ∼

DIEPF2 e

B = T∼ ∼

= {0/0, 2; 1/0, 2; 0, 11/0, 5; 0/1},

= {0/0, 2; 0, 38/0, 2; 1/0, 5; 0, 11/1; 0/1},

DIEPF3

= {0/0, 2; 0, 22/0, 5; 1/1; 0/1} [13].

Method for the Criticality Level Assessment

157

Table 1. The result of a pairwise comparison of the parameters of the criticality level of IPCS Le e\e’ 1 2

3

4

5

6 7

8 9 10

11

12

13

14

15

ωe

e

1

1 1/7 1/6 1/5 1/7 – 1/9 – 1 1/8 1/7 1/7 1/6 1/6 1/8 0,244

0,014

2

7 1

3

8

9

– 1/3 – 7 2

3

2

5

4

2

2,602

0,144

3

6 1/3 1

8

8

– 1/6 – 6 5

6

6

8

7

8

2,934

0,162

4

5 1/8 1/8 1

7

– 1/5 – 5 4

5

5

8

6

8

1,942

0,107

5

7 1/8 1/8 1/7 1

– 1/4 – 7 6

6

7

8

8

4

1,806

0,1

6

– –

–

–

–

– –

– – –

–

–

–

–

–

–

–

7

9 3

6

5

4

– 1

– 9 3

4

5

5

5

3

3,477

0,192

8

– –

–

–

–

– –

– – –

–

–

–

–

–

–

–

9

1 1/7 1/6 1/5 1/7 – 1/9 – 1 1/8 1/7 1/7 1/6 1/6 1/9 0,243

0,013

10

8 1/2 1/5 1/4 1/6 – 1/3 – 8 1

3

3

4

5

3

1,294

0,072

11

7 1/3 1/6 1/5 1/6 – 1/4 – 7 1/3 1

2

3

5

2

0,949

0,052

12

7 1/2 1/6 1/5 1/7 – 1/5 – 7 1/3 1/2 1

3

4

2

0,854

0,047

13

6 1/5 1/8 1/8 1/8 – 1/5 – 6 1/4 1/3 1/3 1

3

3

0,616

0,035

14

6 1/4 1/7 1/6 1/8 – 1/5 – 6 1/5 1/5 1/4 1/3 1

2

0,505

0,028

15

8 1/2 1/8 1/8 1/4 – 1/3 – 9 1/3 1/2 1/2 1/3 1/2 1

0,613

0,034

18,079 1

Note that these are hypothetical intervals and a standard and they have no physical significance. We introduce for this parameter the sensors corresponding to its intervals 3  S14q (tT )} = {S141 (tT ), S142 (tT ), S143 (tT )}. S14 = { q=1

4.2 Results and Discussion We will measure the values of the parameters during a period of 5 min with their fixation every 10 s, i.e. by making 30 measurements. And also we will fix indicators of sensors at the moment of measurement. The results are presented in Table 2 [13, 15, 16]. According to expression (4) the sensor S141 (tT ) is defined as  1, if L14 (tT ) ∈ N141 , T = 1, 30 and obviously at the moments of time S141 (tT ) = 0, if L14 (tT ) ∈ / N141 t23 , t25 , t30 value. S141 (t25 ) = S141 (t30 ) = 1, and at other times, respectively, will take the value 0. Next, determine the counters of the sensors CS14 according to expression (5) and display them in Table 3 [14–16]. Having obtained the values of the counters, we calculate the correction standards of the FN by expression (7). Yes, we will have: T∼ E

141

= CS141 · T∼ e

141

= { 0/0, 6; 1/0, 6; 0, 11/1, 5; 0/3 },

158

A. Gizun et al.

Table 2. The value of parameter L14 (tT ) = LDIEPF (tT ) and their sensors at tT (T = 1, 30) tT L14 (tT ) S141 S142 S143

T∼ E

142 E

T∼

1

39

0

1

0

2

41

0

1

0

3

35

0

1

0

4

56

0

0

1

5

64

0

0

1

6

62

0

0

1

7

50

0

0

1

8

43

0

1

0

9

38

0

1

0

10 35

0

1

0

11 37

0

1

0

12 39

0

1

0

13 41

0

1

0

14 50

0

0

1

15 42

0

1

0

16 54

0

0

1

17 61

0

0

1

18 66

0

0

1

19 51

0

0

1

20 40

0

1

0

21 37

0

1

0

22 25

0

1

0

23 18

1

0

0

24 23

0

1

0

25 19

1

0

0

26 22

0

1

0

27 31

0

1

0

28 35

0

1

0

29 24

0

1

0

30 19

1

0

0

= CS142 · T∼ e

143

142 e

= CS143 · T∼

= { 0/3, 6; 0, 38/3, 6; 1/9; 0, 11/18 },

143

= { 0/1, 8; 0, 22/4, 5; 1/9; 0/9 },

Method for the Criticality Level Assessment Table 3. The frequency of occurrences of the current state of the parameter L14 (tT ) LDIEPF (tT ) (values of sensor counters) CS14

159 =

N14 (i = 1,j = 4,q = 3) N141

CS14q 3

N142 N143 18

9

Similarly count for other parameters. Also the parameters L8 – The ratio of the level of economic losses for the current period to the corresponding level for the previous period, RD and L10 – Specific mortality rate at the moment, RM can be calculated directly by formulas (6) and (7), accordingly LED(t) , where LED(t) – the amount of economic losses for the current period, RD = LED(t−1) SMR(t) LED(t −1)– the amount of economic losses for the previous period and RM = SMR(t−1) , where SMR(t) – mortality for the current period, SMR(t − 1) – mortality for the previous period [13]. The duration of the time intervals used in calculating these parameters must be the same. The indicator of criticality of IPCS is formed based on the values of these parameters, taking into account the ranking procedure, which is presented in Fig. 3. It reflects the level of evaluation parameters and calculates the overall level of criticality. This result confirmed the adequacy of standards developed and correctness of the choice set evaluation parameters. Thus, the result is shown in linguistic form, i.e. a certain level of “Maximum” for the situation caused by the denial of service on the mail server [16–18]. It should be noted that the National Aviation University does not have any system for assessing the criticality of crisis situations of information systems.

Fig. 3. Image of the indicator of the level of criticality of IPCS

160

A. Gizun et al.

5 Conclusions and Future Researches The proposed method based on fuzzy logic methods and expert approaches, allows to assess the criticality of the current situation. The use of expert methods is explained by the need to reduce time and production resources, as the mathematical apparatus of these methods does not require the collection and processing of statistical data. The method consists of 6 stages: determination of criticality assessment parameters, formation of evaluation standards, calculation of CI, measurement and fuzzification of parameters, calculation of criticality level of CS, visualization of results. The results are visualized through an indicator, which is an interface with two panels for displaying the values of the criteria. The values of fuzzy parameters are displayed in the form of a petal diagram, and some parameters that can be estimated using clear data are displayed in the form of a histogram. In addition, the developed indicator of the level of criticality allows to assess the dynamics of the situation, to select effective means and response measures, to facilitate the decision-making process in conditions of uncertainty and the impact of the CS. There is also an example of using the E  method, so the study proposed a set of parameters L = { Le } = {L1 , L2 , ..., LE } = e=1

{T , DVF, GS, OS, OLED, RTLH , F, DDI , CRT , CRP, LM , DIEPF, DVChS}, which is universal and can be used to assess any IPCS regardless of the nature of their origin. Many parameters can be changed by adding or changing the specific criteria to individual IPCSs and CSs depending on the application needs. The concept of the level Ei

(ie ∗ Lie ), of criticality of the situation is introduced based on this set LCSi = e=1

which is determined by the functional relationships between the parameters of the level of criticality. Based on the proposed method, appropriate crisis assessment software was developed and experimental studies were conducted, which confirmed the reliability of theoretical and practical results of the method on the ability to detect and assess crisis situations. In the future, it is planned to develop rules for a wider range of parameters and implement the method in the work of system administrators.

References 1. Gizun, A., Hriha, V., Roshchuk, M., Yevchenko, Y., Hu, Z.: Method of informational and psychological influence evaluation in social networks based on fuzzy logic. In: 4th International Scientific-Practical Conference Problems of Infocommunications Science and Technology, pp. 444–448 (2019) 2. Zahran, B., Al-Azzeh, J., Gizun, A., Griga, V., Bystrova, B.: Developing an expert system method for assessment of destructive information-psychological influence. Indon. J. Electr. Eng. Comput. Sci. 15, 1571–1577 (2019) 3. Gizun, A., Gnatyuk, V., Balyk, N., Falat, P. : Approaches to improve the activity of computer incident response teams. In: Proceedings of the 2015 IEEE 8th International Conference on Intelligent Data Acquisition and Advanced Computing Systems: Technology and Applications, IDAACS 2015, pp. 442–447 (2015)

Method for the Criticality Level Assessment

161

4. Karabutov, N.: Structural identifiability of nonlinear dynamic systems under uncertainty. Int. J. Intell. Syst. Appl. (IJISA) 12(1), 12–22 (2020) 5. Cheng, Q., Yu, L.: Operational mechanism and evaluation system for emergency logistics risks. IJISA 2(2), 25–32 (2010) 6. Hu, Z., Gizun, A., Gnatyuk, V., Kotelianets, V., Zhyrova, T.: Method for rules set forming of cyber incidents extrapolation in network-centric monitoring. In: 2017 4th International Scientific-Practical Conference Problems of Infocommunications Science and Technology (PIC S&T), pp. 444–448. IEEE (2017) 7. Hu, Z., et al.: Statistical techniques for detecting cyberattacks on computer networks based on an analysis of abnormal traffic behavior. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 12(6), 1–13 (2020) 8. Pradhan, P.L.: Application of combinatory mechanism on RTOS UFS ACM for risk optimization. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 8(6), 52–58 (2016) 9. Nasser, A.A., Al-Khulaidi, A., Aljober, M.N.: Measuring the information security maturity of enterprises under uncertainty using fuzzy AHP. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 10(4), 10–25 (2018) 10. Korchenko, A., Kozachok, V., Gizun, A.: Method of criticality level assessment for crisis management systems. Ukr. Inf. Secur. Res. J. 17(1), 86–98 (2015) 11. Mohammadi, F., Bazmara, M., Pouryekta, H.: A new hybrid method for risk management in expert systems. Int. J. Intell. Syst. Appl. (IJISA) 6(7), 60–65 (2014) 12. Coombs, W.T.: Conceptualizing Crisis Communication Handbook of Crisis and Risk Communication, pp. 100–119. Routledge, New York (2009) 13. Fedushko, S., Ustyianovych, T., Gregus, M.: Real-time high-load infrastructure transaction status output prediction using operational intelligence and big data technologies. Electronics 9(4), 668 (2020) 14. Gizun, A.: Methods and means of assessing security parameters to identify crisis situations in the information sphere [dissertation Ph.D.]. National Aviation University (2015) 15. Gordieiev, O., Kharchenko, V.: Profile-oriented assessment of software requirements quality: models, metrics, case study. Int. J. Comput. 19(4), 656–665 (2020). https://doi.org/10.47839/ ijc.19.4.2001 16. Parthasarathi Patra, H., Rajnish, K.: Fuzzy based parametric approach for software effort estimation. Int. J. Mod. Educ. Comput. Sci. (IJMECS) 10(3), 47–54 (2018) 17. Efimov, V., Kotenko, I., Saenko, I.: Network application-layer protocol classification based on fuzzy data and neural network processing. Int. J. Comput. 19(3), 335–346 (2020). https:// doi.org/10.47839/ijc.19.3.1877 18. Chumachenko, D., Sokolov, O., Yakovlev, S.: Fuzzy recurrent mappings in multiagent simulation of population dynamics systems. Int. J. Comput. 19(2), 290–297 (2020). https://doi. org/10.47839/ijc.19.2.1773

Analysis of Key Elements for Modern Contact Center Systems to Improve Quality Bekmurza Aitchanov, Olimzhon Baimuratov(B) , Muratbek Zhussupekov, and Tley Aitchanov Suleyman Demirel University, Kaskelen, Kazakhstan {bekmurza.aitchanov,olimzhon.baimuratov}@sdu.edu.kz, [email protected]

Abstract. Resource management in IT companies with a dynamic change in the market of prices for services, costs, and destabilization of the state, both economic and political, require improvement of the model and management tools. In this paper, it is considered the basic management models, which demonstrate the advantages and disadvantages of resource management models in these areas of IT, also provides more detailed methods for their implementation based on examples and practices of foreign companies. Practical part of this article contains methodology of resource accounting in IT service delivery, fully-fledged tool of analysis, forecasting and informational support for making strategic and tactical management decisions for IT-managers in different levels and provision of reports on effectiveness of each separate services, rational usage of the particular types of resources. For the contact center services forecasting methodology of the service delivery, methodology of workload sharing of the contact center for optimal use of labor resources, infrastructures are presented. It allows keeping balance in distributions of tasks that helps to reduce additional costs and to provide qualitatively monitoring and service. Keywords: Organizational model · Resource management · Resource accounting · Contact center · Organizational processes · Object-oriented organizational model

1 Introduction Resource management of the company today is one of the important areas in the economy, business, and the IT industry. The analysis of planning and resource management models showed that today it is one of the central aspects of the development of the entire industry and providing a guarantee of its sustainability. As a result of the analysis of modern solutions and management methods, there is a big dynamics between organizational processes and mechanisms of operational actions. A comparative analysis of management models makes it possible to more accurately determine the main business goals, criteria, and the final result. The main attention in this paper is given to the company’s IT resource management model to identify key indicators. As an example, the © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 162–174, 2021. https://doi.org/10.1007/978-3-030-80472-5_14

Analysis of Key Elements for Modern Contact Center Systems

163

methodology for calculating the forecasts of the load on the Contact Center is presented, which is a description of the mathematical apparatus for solving the tasks of automating the planning processes and managing the working time of the Contact Center employees to increase the Contact Center productivity by effectively forecasting the load and planning the optimal number of employees. The basic idea is to define optimal combination of infrastructure, operational burden on software and needs of human resources. By analyzing the works in load distribution area, load balancing, management models and methods, technical means and characteristics it was determined that most of existing systems in Contact Centers vary greatly in internal structure of system, service pricing, operation cycle. Lack of common understanding in this direction leads to more expenses, at times governmental and private enterprises cannot adapt during dynamic demand and increasing consumer requirements. Workload forecasting in contact centers with consideration of demographic growth and the needs of some economic sectors doesn’t always provide with positive result, as it’s hard to predict exactly the dynamic of development of competitive environment in 5 or more years, especially in developing countries. The difficulty is that often the data is incomplete and market or the economic sectors in a stage of intensive development, as in or case. Considering the issues and decisions in [1–20], for construction of optimal model of Contact Center system for the Republic of Kazakhstan, a relevant and primary task is to analyze principles of resource management in the Contact Center.

2 Review and Analysis of Planning and Resource Management Models The model of planning and resource management today is one of the central aspects of the development of the industry and provides a guarantee of its sustainability. As a result of the analysis of modern solutions and methods, there is a big dynamics between organizational processes and mechanisms of operational actions. This situation finds a solution to optimization problems. Comparative analysis of models often allows you to more accurately determine the main business goals, criteria, and the final result. As is known from [1–6], the requirements for the entire system determine its basic model. Let’s consider some of them: A large number of commercial workflow management systems demonstrate organizational metamodels that have been developed using a technology-oriented approach. Figure 1 shows the resource model or the IBM FlowMark metamodel, which provides several types of entities with a limited number of elements (for example, without matrix organizations).

Fig. 1. Organizational metamodel staffware 97 [2]

164

B. Aitchanov et al.

Object-Oriented Organizational Model (OMM) aims to separate organizations and roles in the context of e-commerce applications [1], which consists of enterprise, organization, member, and virtual communication object types. Although an enterprise is a collection of several organizations, each organization consists of several members that share common attributes. Member objects are elementary resources that are mapped to the actual resources of the enterprise. Virtual links provide relationships between members of the same or different organizations. Other authors introduced the Organizational Resource Model (ORM), which is an independent repository of organizational structures that were used (terminated) by the WorkParty workflow management system. It provides a semantically rich metamodel that can be modified by inheriting existing entity types. However, some of the provided object types cannot be used for the assignment process, such as resources or their ownership [2]. An organization-oriented approach can be found in environments where workflow management systems are implemented in larger organizations, and a formally defined organizational structure of the enterprise already exists. However, in most commercial workflows in management systems, there are no basic objects defined in organizational theory, such as an organizational position or project team [2]. This conflict can only be resolved if the document management system allows a link to an external storage or if existing types of objects can be changed by the actual structure of the enterprise. Based on the analysis of other models, some authors proposed a general metamodel, shown in Fig. 2, which was developed to meet the needs of both workflow providers who follow the technological approach to resource modeling and suppliers who want to include a complex organizational model in their work management system processes. It can be used to develop metamodels for specific products, but it can also be used to transform resource models from various vendors [2].

Fig. 2. Generic meta model [2]

Analysis of Key Elements for Modern Contact Center Systems

165

The IT function is responsible for a wide range of core activities, such as systems development and support, outsourcing management, strategic planning, project management, technological experiments, improving business processes, and developing IT infrastructure [3].

Fig. 3. IT management model and IT function contribution

The nature and variety of knowledge and skills of IT professionals are directly related to various IT management profiles and directly affect the nature of the contribution of IT functions to the organization. The relationship of an IT function with business units and its organizational environment can take various forms and can vary depending on the main purpose of the IT function [4–6]. The relationship between the profile of the IT management model and the contribution of IT functions to the organization’s work is shown in Fig. 3.

3 Methods of Accounting Resources, Assets, and Processes of It Companies The implementation of the above models requires the development of a methodology for maintaining algorithms of relationships between resource objects, assets, and processes, based on the principles of functional cost accounting (ABC). It defines the basic principles and procedures for organizing accounting in an IT company based on ABC. ABC is a very suitable cost control method for e-business, almost all of its activities are associated with the category of indirect costs [7]. ABC is used to measure costs, the effectiveness of appointments in Scrum companies, which allows companies to switch to a more economical process, which increases their sustainability [8]. The practical implementation of the principles of separate accounting is based on the following approaches. Costs are associated with certain economic resources for which they are designed to ensure operability (for example, salaries of personnel, fuel, and spare parts for vehicles) - the reason for the costs is resources. Economic resources are necessary to carry out certain business processes, perform certain types of work - the reason for the cost of resources is business processes (types of work). Business

166

B. Aitchanov et al.

processes (types of work) in a certain sequence and relationship allow you to produce and sell goods (services) - the reason for these business processes (types of work) are products (services). Thus, costs are distributed between resources, business processes (types of work), and then between services. The basis for the correct distribution is carefully selected indicators - the distribution base of resources between business processes, the distribution base of business processes between services. This approach most accurately reflects the causal relationship of costs with services, allows you to determine the direct costs of resources, business processes, and services, provides transparency in the distribution of costs. Nevertheless, there are certain technological features for IT companies that complicate the above approaches, namely the use of one infrastructure and/or software that can be involved in the provision of many types of services (data center, server, hardware, and software complex contact center, etc.). As well as the availability of various cloud solutions for providing the same services as expected (workflow, management systems, BI systems, etc.). In this regard, it is necessary to supplement these approaches with intermediate stages of cost allocation by type of service (distribution by infrastructure elements, internal services, distribution of areas of activity). The general distribution is carried out in stages, the general logic is presented in Fig. 4. Costs are associated (distributed) with certain economic resources for which they are intended to ensure operability. Economic resources are allocated to the relevant business processes in which they are involved. The costs associated with supporting business processes are distributed between the main business processes and business management processes. The costs associated with the main business processes for maintaining and maintaining the infrastructure/software are allocated to infrastructure/software and related internal services. The costs associated with the main business processes for the implementation (sale) of services and customer service are allocated to the relevant external services. The costs associated with business management processes are allocated to the appropriate infrastructure/software and services, depending on the type. The cost of infrastructure/software, depending on their participation, is shared between the respective internal services. The costs of domestic services, depending on the type of service, are associated with the corresponding external services. Thus, the costs of external services provided to users of services represent the sum of the costs of the corresponding internal services, the costs of business processes associated with the implementation of services, and the distributed costs (auxiliary business processes and business management processes). To allocate costs and determine the degree of involvement of assets, a distribution methodology is used based on a causal relationship of costs and assets with related resources, business processes, network elements, activities, and services. The depreciation of assets involved is allocated based on the distribution of assets involved. The general procedure for the allocation of costs is based on accounting principles based on ABC and by these approaches. For each structural unit/project of the IT company, the corresponding resources (involved assets, personnel, and others) and

Analysis of Key Elements for Modern Contact Center Systems

167

Fig. 4. General scheme of IT company’s cost allocation

costs associated with the content of these resources are allocated. Then, for each structural unit and allocated resources, the corresponding business processes (that is, the processes in which these resources are occupied) are determined. Also, these business processes are correlated with the corresponding elements of infrastructure/software (for activities related to the formation and operation of a data center/hardware and software system/corporate IP), as well as with internal and external services. Data on elements of infrastructure/software, internal and external services are grouped by relevant areas of activity. The study aims to research the impact of knowledge risk management on the effectiveness of an organization, and indicators are considered as “softer” performance indicators, that is, innovativeness, responsiveness, stability, and flexibility [9].

4 Methodology for Calculating Contact Center Load Forecasts The methodology for calculating the forecasts of the load on the Contact Center is a description of the mathematical apparatus used to solve the problems of automating the planning processes and managing the working time of Contact Center personnel. Where the main goal is to increase productivity by effectively forecasting the load on the Contact Center based on available historical information for previous periods, as well as planning the optimal number of staff. The objective of this Methodology is to calculate the shift of agents depending on the key parameters of the Contact Center: the number of agents, queues, and degree of load, control of the volume of calls processed, and accuracy of adherence to the schedule in real-time. The models presented in Fig. 5 are used as the mathematical base of this Methodology. To predict the load on the network of the Contact Center, taking into account the analysis of seasonal, weekly, and daily fluctuations, a forecast using the method of exponential smoothing with the trend and seasonality of Holt-Winters is used [10].

168

B. Aitchanov et al.

To calculate the number of operators, depending on the number of calls processed at the Contact Center for a certain period, it is used the Erlang C load calculation model using data on the predicted load [11]. The smoothing coefficient in the Holt-Winters model is “wired” into the algorithm and is assumed to be 0.2.

Fig. 5. Contact center working time planning algorithm

A. Description of the Work Process of the Contact Center For clarity of understanding, we present a definition of the term: Process - a stable and targeted set of interrelated actions that, according to a certain technology, transform the inputs into outputs to obtain predetermined products, results, or services that are of value to the consumer.

Analysis of Key Elements for Modern Contact Center Systems

169

The change in time is a characteristic feature of the contact center: the number of working agents changes, the spread of the queue length, and the degree of load of the operators is characteristic, the volume of processed calls is unstable and the schedule is observed in real-time. Thus, the contact center functioning process is a set of values of the above values distributed over time, which is continuously discrete in nature (some of the values are distributed continuously, and some are discrete), thereby representing a fairly common type of distribution - a time series. Events that make up the time series can be analyzed according to the rank of the magnitude of the event and according to the sequence of events. When ranking events, the order of their appearance is neglected, while studying the sequence of occurrence of events, on the contrary, it is assumed that past events from a given time series can affect the magnitude of current or future events [10, 12]. The time series of the contact center functioning processes [12] contains a systematic component (usually including several components) and random noise (error), which makes it difficult to detect regular components. As a rule, methods for studying time series include various noise filtering methods [10], which make it possible to see the regular component more clearly. In [13], a description in the discrete-time domain is presented and an analysis of system components is performed. One of the particular issues considered in this paper is the analysis of the components described by various types of equations. The results of our analysis are presented in the form of stochastic Volterra models that can be used both for analysis and for the synthesis of control systems. Traditional methods of data analysis (statistical methods), OLAP, and traditional visualization (reports, dashboards) are mainly focused on testing pre-formulated hypotheses (verification-driven data mining) and on “crude” exploratory analysis, which forms the basis of operational analytical data processing (OLAP) [14]. Moreover, while OLAP and traditional visualization are currently rarely used for time series analysis, the apparatus of statistical mathematics is still the main one, at least when analyzing the work of the Contact Center. It is the tools of statistical mathematics that are the Holt-Winters and Erlang C. models. The analysis method proposed in this technique is distinguished by the creation of some Data Mining add-in that takes into account the importance of the influence of individual characteristics of operators (project ownership, knowledge of languages) [15–18]. B. Forecasting the Load on the Contact Center When smoothing the time series, the irregular response component has a greater or lesser effect, so the smoothed series turns out to be (within the classical model) a superposition of the trend and cyclical (and possibly seasonal) components of the process, which facilitates their further investigation. Commonly used is the moving average method or the method of exponential smoothing; both methods are somewhat subjective for the choice of smoothing parameters. With a constant acceleration of the indicator growth process in a given section of the time series, it makes sense to try to approximate the initial data by an exponential function using the Holt-Winters method for exponential smoothing of the time series [10].

170

B. Aitchanov et al.

Exponential smoothing provides a visual representation of the trend and allows you to make short-term forecasts, and when you try to extend the forecast to a longer period, completely meaningless values are obtained: it seems that the development of the process in the direction of growth or decrease has completely stopped – the same values are predicted for any period of the future response. The more sophisticated Holt-Winters method successfully copes with both mediumterm and long-term forecasts, since it is capable of detecting microtrends (trends related to short periods) at times immediately preceding the forecasting and extrapolating these trends for the future. And although only linear extrapolation to the future is possible, in most real situations it is enough. When using the method, it is necessary to sequentially calculate the smoothed values of the series and the trend value accumulated at any point in the series. The Holt-Winters model consists of the following elements [15–18]: 1. The exponentially smoothed series is calculated: Lt = k ∗ Yt /St−s + (1 − k) ∗ (Lt−1 + Tt−1 )

(1)

where Lt – smoothed value for the current period; k – row smoothing coefficient; St−s – seasonality coefficient of the previous period; Yt – the current value of the series; Lt−1 – smoothed value for the previous period; Tt−1 – trend value for the previous period. For the first period at the beginning of the data, the exponentially smoothed series is equal to the first value of the series. Seasonality in the first and second periods is 1. 2. The trend value is determined: Tt = b ∗ (Lt − Lt−1 ) + (1 − b) ∗ Tt−1

(2)

where T – trend value for the current period; b – trend smoothing coefficient. The trend value for the first period is 0. 3. Seasonality is estimated: St = q ∗ Yt /Lt + (1 − q) ∗ St−s

(3)

where St – seasonality factor for the current period; q – seasonality smoothing coefficient. 4. The forecast is made: Yˆ t+p = (Lt + p ∗ Tt ) ∗ St−s+P

(4)

where Yˆ t+p – Holt-Winters forecast p periods ahead; p – serial number of the period for which the forecast is made; St−s+P – seasonality factor for the same period in the last season. To find the values of the “smoothing coefficients” of the model, the Least Squares Method (LSM) is used—a mathematical method used to solve various problems, based

Analysis of Key Elements for Modern Contact Center Systems

171

on minimizing the sum of the squared deviations of some functions from the desired variables. LSM is one of the basic regression analysis methods for estimating the unknown parameters of regression models from sample data. Let be x a set of unknown variables (parameters), f (x) a set of functions of this set of variables. The task is to select such x values so that the values of these functions are as close as possible to some y values. Essentially, it is talking about the “solution” of an overdetermined system of equations in the indicated sense of the maximum proximity of the left and right sides of the system. The essence of the least-squares method is to choose the sum of the squares of the deviations of the left and right parts as a “measure of proximity”. Thus, the essence of LSM can be expressed as follows:   ei2 = (5) (yi − fi (x))2 → min i

i

x

If the system of equations has a solution, then the minimum value of the sum of squares will be zero and exact solutions of the system of equations can be found analytically or, for example, by various numerical optimization methods. If the system is overdetermined, that is, loosely speaking, the number of independent equations is greater than the number of variables sought, then the system does not have an exact solution and the least squares method allows you to find some “optimal” vector in the sense of the maximum proximity of the vectors.

5 Results and Discussion As a result of the analysis of the main models in [1–6], which demonstrate the main advantages and disadvantages, as well as the direction of application of these models for resource management in IT companies. Also, when analyzing the literature, the results of big data processing in the field of organization management were presented, which are identical in structure to IT companies, as well as companies as an IT industry as outsourcing. From the presented models in the figures above, it can be determined that the structure of the company is closely connected with the main management model, which forms operational characteristics, internal relations, and strategic development. Analysis of resource management models at the enterprise allows us to evaluate both qualitative and quantitative management indicators when creating universal models based on dynamic indicators of the system. From an analysis of previous works it was found that one of the main parameters is adaptability, stability, and scalability. One of the important features of the formation of a resource management model is the structure of the organization and its infrastructure part, as well as a direct dependence on the use of external resources. No less important is the internal management policy and information security policy at the enterprise, as well as the quality and format requirements for the provision of services and products to a client or user. It is also necessary to note the very dynamics in management and in the mechanism due to changes in business processes that determine strategic plans and goals [21]. These features of resource management model formation and technological features for IT-company in the use of one infrastructure and software for realization of different

172

B. Aitchanov et al.

types of IT-services were solved within the proposed in the work modified generic model Activity-based costing (ABC) for an IT-company. Intermediate stages of cost distribution by types of services, infrastructural elements, record of internal service, fields of work and business process were imposed. These stages were formalized as a versatile step-by-step framework, which can be used to record different types of IT-services [22]. The developed accounting system is a full-fledged analysis, forecasting and information support tool for making strategic and tactical management decisions for IT managers at various levels and reporting on the effectiveness of each service, providing a clear separation of income and expenses between different types of activities, services, businessprocesses. Efficiency (profitability) for the company of each client or an arbitrary group of clients (consumer categories) in the context of all types of services. Rational use of certain types of resources. For the planning of the IT service delivery resources, in this work, forecasting of the service delivery level of the Contact Center based on the expected flow of incoming requests is discussed. The working hours planning algorithm of the Contact Center on the basis of existing historical information for the preceding period was introduced. The task of improving implementation of labor resources for service delivery is discussed by planning the optimal staff numbers. For the solution of this task the methods of statistical mathematics, especially models of Holt-Winters and Erlang C. are proposed. The ways of analysis, that offered in this methodology differs with the creation of a superstructure Data Mining, adapted to the importance of the impact of individual characteristics of operators, such as project ownership, language skills. It allows to consider the microtrends in an instant of time, immediately preceding forecasting, and to extrapolate this trends for the future, to make a medium-term and long-term forecasts. The methodology for calculating the forecasts of the load on the Contact Center is the practical basis for the implementation of the resource management model in the IT company: Contact Center.

6 Conclusions The main idea of this work is to analyze existing models, methodology and principles of management in it companies, where aspects such as business strategy, it management and business are integrated under management. Maintaining a management balance in the presence of many interference factors exposing to the manifestation of risks is a difficult task that requires the development or improvement of a management model. Basic methods based on the theory of control systems, strategic planning, etc. To implement the resource management model in an IT company, the results of this analysis determine the criteria and modern methods based on the practice of identical foreign companies. Currently, in the Kazakhstan ecosystem of an IT company, there is no single, generalized model that is adaptive to risk and interference factors, often leading to significant damage. The use of a service-oriented approach in some cases contributes to the creation or development of a new IT segment. As an example, we consider the task of distributing the load on the contact center with the classic control model to determine the difference between the existing models in Kazakhstan, which in the first place must be adaptive and stable both external and internal interference factors. Further study of

Analysis of Key Elements for Modern Contact Center Systems

173

this problem and improvement of Contact Center system model adaptive for economic sectors of Kazakhstan are expected.

References 1. Cheng, E.: An object-oriented organizational model to support dynamic role-based access control in electronic commerce applications. In: Proceedings of the 32nd Hawaii International Conference on System Sciences – 1999 (1999) 2. Muehlen, M.: Resource modeling in workflow applications. In: Proceedings of the 1999 Workflow Management Conference (WFM99), pp. 137–153 (1999) 3. Wiedemann, A., Gewald, H.: Examining cross-domain alignment: the correlation of business strategy, IT management, and IT business value. Int. J. IT/Bus. Align. Govern. 17–18 (2017) 4. Walter, R.: Organization and role models for workflow processes. In: The Workflow Management Coalition Handbook, Winchester, pp. 165–172 (1997) 5. Paréa, G., Guillemetteb, M.G., Raymondc, L.: IT centrality, IT management model, and contribution of the IT function to organizational performance: a study in Canadian hospitals. Inf. Manage. 57(30), 103198 (2020) 6. Buyya, R., Abramson, D., Gid, J., Stockinger, H.: Economic models for resource management and scheduling in Grid computing. Concurr. Comput.: Pract. Exp. 14, 1507–1542 (2002) 7. Iltuzer, Z., Tas, O., Gozlu, S.: Implementation of activity-based costing in e-businesses. In: PICMET 2007 Proceedings, pp. 1119–1125 (2007) 8. Kom, E.A.P., Suryani, E.: Designing cost measurement system in a small scrum based software company using activity based costing model (case study: ABC company). In: 2019 International Conference on Information and Communications Technology (ICOIACT), pp. 943–947 (2019) 9. Durst, S., Hinteregger, C., Zieba, M.: The linkage between knowledge risk management and organizational performance. J. Bus. Res. 105, 1–10 (2019) 10. Hyndman, R.J., Athanasopoulos, G.: Forecasting: Principles and Practice, 2 edn., pp. 198– 203. OTexts (2018) 11. Archawaporn, L., Wongseree, W.: Erlang C model for evaluate incoming call uncertainty in automotive call centers. In: 2013 International Computer Science and Engineering Conference (ICSEC), pp. 109–113 (2013) 12. Motta, G., Barroero, T., Sacco, D., You, L.: Forecasting in multi-skill call centers: a multiagent multi-service (MAMS) approach: research in progress. In: 2013 Fifth International Conference on Service Science and Innovation, pp. 223–229 (2013) 13. Aitchanov, B.H., Nikulin, V.V., Baimuratov, O.A.: Mathematical modeling of digital pulsefrequency modulation control systems developed for objects with transport delay. In: 2013 25th Chinese Control and Decision Conference (CCDC), pp. 1407–1411 (2013) 14. Sumathi, S., Sivanandam, S.N.: Introduction to Data Mining and its Applications. Springer, Heidelberg (2006). https://doi.org/10.1007/978-3-540-34351-6 15. Rusavska, V., Melikh, T., Melikh, O., Chala, T., Babushko, S., Halytska, M.: Estimation of the influence of the seasonality factor in the strategic activity of tourism and hospitality enterprises. Int. J. Innov. Technol. Explor. Eng. (IJITEE) 9(1) (2019) 16. Ding, S., Koole, G., van der Mei, R.D.: On the estimation of the true demand in call centers with redials and reconnects. Eur. J. Oper. Res. 246(1), 250–262 (2015). https://doi.org/10. 1016/j.ejor.2015.04.018 17. Pyon, C.U., Woo, J.Y., Park, S.C.: Intelligent service quality management system based on analysis and forecast of VOC. Expert Syst. Appl. 37(2), 1056–1064 (2010). https://doi.org/ 10.1016/j.eswa.2009.06.066

174

B. Aitchanov et al.

18. Chromy, E., Misuth, T., Weber, A.: Application of Erlang formulae in next generation networks. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 4(1), 59 (2012) 19. Alguliyev, R.M., Nabibayova, G.Ch., Gurbanova, A.M.: Development of a decision support system with the use of OLAP-technologies in the national terminological information environment. Mod. Educ. Comput. Sci. (6), 43–52 (2019). https://doi.org/10.5815/ijmecs.2019. 06.05 20. Al-Samawi, Y.: Digital firm: requirements, recommendations, and evaluation the success in digitization. Inf. Technol. Comput. Sci. 1, 39–49 (2019). https://doi.org/10.5815/ijitcs.2019. 01.05 21. Tikhomirov, A., et al.: Network society: aggregate topological models. In: Dudin, A., Nazarov, A., Yakupov, R., Gortsev, A. (eds.) ITMM 2014. CCIS, vol. 487, pp. 415–421. Springer, Cham (2014). https://doi.org/10.1007/978-3-319-13671-4_47 22. Sklyar, V., Kharchenko, V.: Assurance case for safety and security implementation: a survey of applications. Int. J. Comput. 19(4), 610–619 (2020). https://doi.org/10.47839/ijc.19.4.1995

End-to-End Scene Text Recognition Network with Adaptable Text Rectification Yi Zhang1 , Zhiwen Li2 , Lei Guo2 , and Wenbi Rao1(B) 1 School of Computer Science and Technology, Wuhan

University of Technology, Wuhan 430070, China [email protected] 2 PipeChina West East Gas Pipeline Company, Wuhan 430074, China

Abstract. Scene text recognition has attracted wide attention of academic, since its irregular shape makes text recognition difficult. Because of the influence of angle, shape and lighting, processing perspective text and curved text still faces various problems. This paper presents an end-to-end scene text recognition network (SRATR). SRATR consists of a rectification network based on spatial transform network and an attention-based sequence recognition network. The rectification network is responsible for rectifying the irregular text, which plays a significant role in text recognition network. In addition, the training needs only scene text images and word-level annotations. The recognition network uses the encoderdecoder mechanism to extract feature sequence from the rectified text. Then we translate the feature sequence into a character sequence to output. In the decoder part, we proposed a fractional pickup method, which can eliminate the interference of noise from the text, make the decoder generate a correct region of focus and improve accuracy of text recognition. This is an end-to-end recognition network. Experiments over several of public datasets prove that SRATR has an outstanding performance in recognizing irregular text. Keywords: Scene text recognition · Text rectification · Perspective text and curved text

1 Instruction Scene text recognition is essential in the field of text recognition. The text in the natural scene image contains rich and accurate high-level semantic information, which has guiding significance for blind navigation, intelligent city traffic management, driverless driving of cars, and instant translation. The previously proposed Optical Character Recognition (OCR) has been able to process document text effectively; it is still a big challenge for scene text recognition. The text in the scene image as shown in Fig. 1, which is perspective text and curved text, has various difficulties to recognition. To enhance the recognition efficiency of scene text, we introduce an end-to-end Scene text Recognition network with Adaptable Text Rectification (SRATR). The SRATR includes a text rectification network and a text recognition network. Text rectification © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 175–184, 2021. https://doi.org/10.1007/978-3-030-80472-5_15

176

Y. Zhang et al.

Fig. 1. Some pictures with curved text and perspective text

network rectifies irregular text to make it easier to recognize and text recognition network recognizes rectified text using sequence recognition method. We divide the text recognition task into two steps. The text rectification network uses the Spatial Transformation Network (STN) to process the input image with curved and perspective text [1]. Firstly, the localization network locates the irregular text in the input image and generates control points. Secondly, the grid generator is used to estimate the parameters of the Thin Plate Spline (TPS) transformation to generate a sampling network. Finally, the rectified image that generated through the sampling network becomes horizontal, tightly-bounded, and easier to recognize. The text recognition network recognizes the rectified image and predicts a character sequence. We extracted the feature sequence of rectified image by the encoder firstly. To obtain the character sequence, we use the attention-based decoder to process the extracted feature sequence. In the decoding process, the fractional pickup method, which is mainly applied to text surrounded by noise, is proposed. The fractional pickup method can obtain features of the neighboring region and improve the recognition effect in training. It is an end-to-end text recognition network.

Fig. 2. Structure of the whole network

2 Related Work Scene text detection and recognition has received extensive attention in recent years, so a wealth of papers on scene text recognition have been published. Related surveys are in references [2, 3].

End-To-End Scene Text Recognition Network

177

Scene text detection is a crucial stage of scene text recognition. However we use text rectification instead of text detection to locate and rectify text [4]. In the early days, the rectification of irregular texts mainly based on morphological analysis or registration techniques on document [5]. It is mainly used in image rectification that contains multiple lines of text. However, scene text images often appear in the form of single words, Bartz solved problem through the integrated rectification-recognition network [6]. Zhan proposed a method for rectifying irregular text more effectively [7]. Initially, the research of text recognition was based on the document text. Then, scene text recognition method based on deep neural networks becomes prevalent [8]. Jaderberg proposed a convolutional neural network method for unconstrained recognition [9]. With the wide application of convolutional neural networks, Shi proposed an end-toend network with CNN and RNN that can better learn contextual information [10]. Zhang adapted the sequence-to-sequence domain to the network for robust text image recognition [11]. The end-to-end recognition problem has been extensively studied. Jaderberg proposed an end-to-end system for text spotting and improved it [12]. TextBoxes detects and recognizes text in an end-to-end mechanism, achieving a competitive result [13]. Zhan proposed an improved method for text rectification [7]. This method uses multiple iterations of TPS transform to rectify text. Our method is mainly implemented by a rectification-recognition model which is trained in an end-to-end way [14].

3 Methodology 3.1 Text Rectification Network This part uses STN proposed by Jaderberg. It is differentiable and can back propagate the error differential. In addition, we use TPS as a 2D transformation method, which can flexibly deform the image. Figure 2 is the framework of the whole network. The first step of STN is to generate two sets of control points K through the localization network. As shown in Fig. 3, these two sets of control points are distributed on the upper and lower sides of the text at the same interval, The regression of the control points C ’ is determined by x, y coordinates. We use a standardized coordinate system. It puts the center point of the input image as the origin, so the coordinates of the top left corner are (−1, −1), and the bottom right corner are (1, 1). The coordinates of the input image I are defined as  T   C  = C1 , C2 , . . . , Ck ∈ R2×k , and Ck = xk , yk’ is the x, y coordinates of the k-th point. The coordinates of the rectified image Ir are defined as C = [C1 , C2 , . . . , Ck ]. We use CNN in the regression process. The second step of STN is that the grid generator estimates the parameters of the TPS transform and generates a sampling grid. As described in Fig. 3, we define C on the rectified image Ir as the basic control points, and C ’ on the input image I correspond to C on a one-to-one  Any point (xi , yi ) on Ir , the TPS transform can find the corre basis. sponding point xi , yi on I. The parameters of the TPS transformation are represented by a matrix T, and the calculation method is:   (1) T = C  02×3 −1 C

178

Y. Zhang et al.

To find the corresponding point pi on I, we use TPS transformation. It takes C and  p as input and outputs p . By iterating all points in P, we can generate the grid P  = pi on the input image I. By the way, all parts in this model are differentiable.

Fig. 3. Processing mechanism of text rectification network.

The last step of STN is to get the output image Ir through the sampler. The pixel value of pi is obtained by performing bilinear interpolation on pi ’ on the input image. After getting all the pixel values, we get the rectified image Ir . Since pi ’ may fall outside the image, the value cropping operation needs to be performed before sampling. The formula for generating the rectified image Ir : Ir = V(P, I)

(2)

Where V represents bilinear interpolation. TPS transformation is very flexible. As shown in Fig. 4, it can rectify irregular text and regular text, to obtain a rectified image. This will be of great help to the readability of the recognition network.

Fig. 4. The rectified result of irregular images.

3.2 Text Recognition Network Recognizing the rectified image Ir that has been obtained and predicting a character sequence are the process to be performed by the recognition network. We model the recognition problem as a sequence recognition problem and then process it.

End-To-End Scene Text Recognition Network

179

3.2.1 Attention-Based Sequence-to-Sequence Recognition Network This part consists of a convolutional layer and a recursive network. We use an attentionbased decoder, which is a unidirectional recurrent neural network [15]. It can access the output of the encoder at each decoding step and can accurately target and label it. The iterative maximum steps of the decoder are T, and finally generate sequence (y1 , y2 , . . . , yT ) directly. When the end-of-sequence symbol (EOS) is predicted, the decoder stops processing. At step t, the output character yt is: yt = Soft max(Wout st + bout ) Where st is the hidden state of t-th step. We use GRU to update st :   st = GRU yprev , gt , st−1

(3)

(4)

In addition, αt,i is a vector of attention weights: L       αt,i = exp et,i / exp et,j

(5)

j=1

Because yt−1 is included in the calculation, the decoder can capture the dependencies between the output characters. In this way, the context information of the previous step can be retained during the training process, making the recognition more accurate. Finally, the decoder outputs the target sequence as the final result. 3.2.2 Fractional Pickup The ability that decoder selects regions of interest is enhanced by properly aligned feedback. The scene text is surrounded by various noises due to the acquisition conditions, which will lead that decoder focus on the wrong background region. This will cause the decoder to generate incorrect regions, select unrelated features, and finally cause false predictions. To solve this problem, we propose a training method called fractional pickup (FP), which partially selects adjacent features during the training process. A wider attention region can improve the robustness of the entire model. We use FP at every step of the decoder to modify the attention weight. At step t, we update the attention weight αt,k and αt,k by the following formula:  = βα αt,k t,k + (1 − β)αt,k+1 (6)  αt,k+1 = (1 − β)αt,k + βαt,k+1 Where β and interger k is generated randomly. We use FP during the training process has the following advantages. First, randomize αt,k and αt,k+1 , which not only makes the distribution of αt different at each step, even for the same image, but also prevents overfitting. Secondly, in Fig. 5, the BLSTM, the output of feature sequence hi , at the training of (k + 1)-th step, will have a shortcut to step k. This shortcut can save some of the features of the previous step, which provide more context information for the subsequent training and improve the robustness. Finally, the changes of the adjacent αt,k

180

Y. Zhang et al.

and αt,k+1 affect the decoder, whereas αt,k+1 and hk+1 are related, so hk will be affected by the neighboring features. We apply the model to the context when extracting features by adding the FP method during the training process. It plays an important role in predicting target characters correctly.

Fig. 5. The fractional pickup mechanism have a shortcut, the green arrow in the figure, connect to step k when training step k + 1.

3.3 Training The training of our model is end-to-end and the parameters are randomized. The training set is defined as D = {Ii , Yi }, i = 1, . . . , N . We minimize the negative log-likelihood of the conditional probability of D as follows: Loss = −

|Yi | N  

  logp Yi,t |Ii ; θ

(7)

i=1 t=1

Where Yi,t is the groundtruth of the t-th character in Ii . θ is defined as the parameter of the model.During the training process, the weights of all layers are randomly initialized except for the localization network. The output fully-connected layer of the localization network is initialized with a weight of 0. Experiments prove that this will make the training process more stable.

4 Experiments We conducted extensive experiments on a list of standard datasets. These datasets include both regular text datasets and irregular text datasets. The performance is measured by the case-insensitive word accuracy. 4.1 Datasets and Experimental Settings Our experiment contains two training datasets and four test datasets. Synth90k is a synthetic dataset generated from 90k English words. It contains 9 million images, all used for training. SynthText is also a synthetic dataset, which generated in a similar way

End-To-End Scene Text Recognition Network

181

to Synth90k [16]. But it is for target detection. IIIT5k-Words (IIIT5k) contains 3000 cropped test images and Street View Text (SVT) contains 647-cropped images from Google Street View [17, 18]. SVT-Perspective (SVT-P) contains 639-cropped images [19]. Since the images are selected from the side-view image of Google Street View, almost of them are severely distorted by non-frontal view angle. This dataset is to evaluate the performance of recognizing perspective text. CUTE80 (CUTE) contains 288cropped images, all of which are from 80 high-resolution images [20]. It is to evaluate the performance of recognizing curve text. The rectification network is composed of 6 convolutional layers, the core size is 3×3 and the number of output filters is 32, 64, 128, 256, 256, 256, 5 2 × 2 max-pooling layers and 2 fully-connected layers. We set the control point to 2K, and K to 20. Finally, the size of the output image by the rectification network is 32 × 100 which the input image size of the recognition network is. The encoder consists of a 45-layer residual network and a 2-layer BLSTM. It is case insensitive, so the decoder normalizes the output to 26 lowercase letters, 10 digits, and an EOS symbol. In the training process, we first set the learning rate to 1.0, and then gradually decay to 0.1, 0.01. Setting the learning rate in this way helps to converge better. We implement the proposed model under the framework of PyTorch training with NVIDIA GTX-1080Ti GPU. 4.2 Experimental Results 4.2.1 Recognizing Irregular Text In order to evaluate the recognition effect of perspective text and curved text, we conducted comparative experiments on SVT-P and CUTE datasets. We made some improvement on the basis of baseline. As can be seen from Table 1, the performance of our method has been improved. Among them, the recognition accuracy on SVT-P is improved by 0.5% compared with baseline, but CUTE is not improved, which shows that our method does not have an ideal performance on curve text. Table 1. The accuracy of text recognition on two irregular text datasets SVT-P and CUTE. Method

SVT-P CUTE

Shi [21]

71.8

Liu [22]

73.5

59.2 –

Cheng [23] 71.5

63.9

Cheng [24] 73.0

76.8

Luo [25]

76.1

77.4

Baseline

77.4

78.8

Ours

77.9

78.8

Figure 6 is the rectified image and recognition result generated by our method. It can be seen that the prediction results of the first three input images are the same

182

Y. Zhang et al.

as their groundtruth, indicating that the prediction are correct (the text in blue color). Then the four input images are fuzzy and the text structure is too complicated, which leads to wrong prediction (the text in red color). For example, the recognition result of “enterprise” is “everprise”, and the recognition result of “grandstand” is “chandstand”. Therefore, we need to study to get a model with higher recognition accuracy.

Input Images

Rectified Images Groundtruth Prediction

meant meant

virgin virgin

imperial imperial

enterprise grandstand manchester for everprise chandstand

Fig. 6. The rectified image and recognition result of regular text generated by our method.

4.2.2 Results on General Benchmarks We conducted experiments on four datasets IIIT5k, SVT, SVT-P and CUTE. These datasets contain some regular text images and images with various types of perspective text and curved text. Firstly, we compare the methods which add fractional pickup with other methods. The FP is mainly to mitigate various types of noise around the text, and the images in the SVT and SVT-P datasets contain a lot of noise and low resolution, so these two datasets can be used for comparison. Table 2. The result of scene text recognition over the datasets IIIT5k, SVT, SVT-P, CUTE. Method

IIIT5k SVT SVT-P CUTE

Jaderberg [9]

–

71.7

–

–

Jaderberg [12]

–

80.7

–

–

Shi [21]

81.9

81.9 71.8

59.2

Liu [22]

83.3

83.6 73.5

–

Cheng [23]

83.7

82.2 71.5

63.9

Cheng [24]

87.0

82.8 73.0

76.8

Luo [25]

91.2

88.3 76.1

77.4

Baseline

92.7

88.0 77.4

78.8

Ours

93.2

89.2 77.9

78.8

It can be concluded from Table 2 that on all datasets, the recognition accuracy of the models containing the FP method exceeds the models without the FP, especially the SVT

End-To-End Scene Text Recognition Network

183

and SVT-P datasets. On the SVT dataset, the recognition accuracy is improved by 1.2%, and on the SVT-P is 0.5%. This shows the effectiveness of the FP method in mitigating the influence of noise in the image and reflects its advantages. Then, we can see that our method has higher recognition accuracy than other methods in the table, which reflects the advantages of our method.

5 Conclusion In this paper, we propose an end-to-end scene text recognition network (SRATR) which includes two parts: text rectification network and text recognition network. In short, the contribution is divided into three parts. We propose the SRATR framework to recognize irregular text first. The image rectified by the text rectification network is easier to recognize. It improves the accuracy of text recognition. Then, We propose a fractional pickup method for the training of attention-based decoder. It expands the visual domain, reduces noise disturbance and further improves the stability of the attention-based decoder. Last, we use back propagation gradient for training throughout the network, which speeds up the convergence. We train our model end-to-end, which only requires images and related text labels. The range of text detection is limited to the vicinity of the target text, and cannot expanded to the whole image. The ability to recognize arbitrary-oriented text is poor. We will work in this direction to solve the above deficiencies. Acknowledgment. This work was supported by National Natural Science Foundation of China (No. 61703316), Funding Project of Postgraduate Joint Training Base of WHUT and CSEPDI.

References 1. Jaderberg, M., Simonyan, K., Zisserman, A., Kavukcuoglu, K.: Spatial transformer networks. In: NIPS, pp. 2017–2025 (2015) 2. Ye, Q., Doermann, D.: Text detection and recognition in imagery: a survey. IEEE Trans. Pattern Anal. Mach. Intell. 37(7), 1480–1500 (2015) 3. Xiyan, L., Gaofeng, M., Chunhong, P.: Scene text detection and recognition with advances in deep learning: a survey. Int. J. Doc. Anal. Recogn. 22(2), 143–162 (2019) 4. Singh, P., Budhiraja, S.: Offline handwritten gurmukhi numeral recognition using wavelet transforms. Int. J. Mod. Educ. Comput. Sci. 4(8), 34–39 (2012) 5. Lu, S., Chen, B.M., Ko, C.C.: Perspective rectification of document images using fuzzy set and morphological operations. Image Vis. Comput. 23(5), 541–553 (2005) 6. Bartz, C., Yang, H., Meinel, C.: SEE: towards semi-supervised end-to-end scene text recognition. In: Thirty-Second AAAI Conference on Artificial Intelligence, pp. 6674–6681 (2018) 7. Zhan, F., Lu, S.: ESIR: end-to-end scene text recognition via iterative image rectification. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 2059–2068 (2019) 8. Mohamed, M.: Smart warehouse management using hybrid architecture of neural network with barcode reader 1D / 2D vision technology. Int. J. Intell. Syst. Appl. 11(11), 16–24 (2019) 9. Jaderberg, M., Simonyan, K., Vedaldi, A., Zisserman, A.: Deep structured output learning for unconstrained text recognition. In: ICLR, pp. 1130–1150 (2015)

184

Y. Zhang et al.

10. Baoguang, S., Xiang, B., Cong, Y.: An end-to-end trainable neural network for image-based sequence recognition and its application to scene text recognition. IEEE Trans. Pattern Anal. Mach. Intell 39(11), 2298–2304 (2017) 11. Zhang, Y., Nie, S., Liu, W., Xu, X., Zhang, D., Shen, H.T.: Sequenceto-sequence domain adaptation network for robust text image recognition. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 2740–2749 (2019) 12. Jaderberg, M., Simonyan, K., Vedaldi, A., et al.: Reading text in the wild with convolutional neural networks. Int. J. Comput. Vision 116(1), 1–20 (2016) 13. Liao, M., Shi, B., Bai, X., Wang, X., Liu, W.: Textboxes: a fast text detector with a single deep neural network. In: Thirty-First AAAI Conference on Artificial Intelligence, pp. 4161–4167 (2017) 14. Ahmed, A.U., Masum, T.M., Rahman, M.M.: Design of an automated secure garage system using license plate recognition technique. Int. J. Intell. Syst. Appl. 6(2), 22–28 (2014) 15. Chorowski, J.K., Bahdanau, D., Serdyuk, D., Cho, K., Bengio, Y.: Attention-based models for speech recognition. In: NIPS, pp. 577–585 (2015) 16. Gupta, A., Vedaldi, A., Zisserman, A.: Synthetic data for text localization in natural images. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 2315–2324 (2016) 17. Mishra, A., Alahari, K., Jawahar, C.: Top-down and bottom-up cues for scene text recognition. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 2687–2694 (2012) 18. Wang, K., Babenko, B., Belongie, S.: End-to-end scene text recognition. In: ICCV, pp. 1457– 1464 (2011) 19. Phan, T.Q., Shivakumara, P., Tian, S., Tan, C.L.: Recognizing text with perspective distortion in natural scenes. In: IEEE International Conference on Computer Vision, pp. 569–576 (2013) 20. Risnumawan, A., Shivakumara, P., Chan, C.S., Tan, C.L.: A robust arbitrary text detection system for natural scene images. Expert Syst. Appl. 41(18), 8027–8048 (2014) 21. Shi, B., Wang, X., Lyu, P., Yao, C., Bai, X.: Robust scene text recognition with automatic rectification. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 4168– 4176 (2016) 22. Liu, W., Chen, C, Wong, K.Y.K., Su, Z., Han. J.: STAR-net: a spatial attention residue network for scene text recognition. In: British Machine Vision Conference, vol. 43 1–13 (2016) 23. Cheng, Z., Bai, F., Xu, Y., Zheng, G., Pu, S., Zhou, S.: Focusing attention: Towards accurate text recognition in natural images. In: IEEE International Conference on Computer Vision, pp. 5076–5084 (2017) 24. Cheng, Z., Xu, Y., Bai, F., Niu, Y., Pu, S., Zhou, S.: AON: towards arbitrarily-oriented text recognition. In: IEEE Conference on Computer Vision and Pattern Recognition, pp. 5571– 5579 (2018) 25. Canjie, L., Lianwen, J., Zenghui, S.: Moran: a multi-object rectified attention network for scene text recognition. Pattern Recogn. 90, 109–118 (2019)

Perfection of Computer Algorithms and Methods

Forced Oscillations with Delays in Elastic and Damping Forces Alishir A. Alifov(B) Mechanical Engineering Research Institute of the Russian Academy of Sciences, Moscow 101990, Russia

Abstract. Forced oscillations in a system with a source of energy of limited power in the presence of delays in the forces of elasticity and damping are considered. The elastic force also includes a non-linear component. The solution of the system of equations with nonlinearity was performed using the method of direct linearization, including the linearization accuracy parameter, and the method of change of variables with averaging built on its basis for solving the linearized equation. The latter method gives a standard form equation for determining the non-stationary and stationary values of the amplitude and phase of the oscillations. Using it and the averaging procedure for solving the energy source equation, the necessary relations for the considered system, as well as the source velocity, are obtained. The stability of stationary motions is considered and stability criteria are derived. Calculations are performed to obtain information about the effect of delays on the dynamics of the system. Keywords: Forced oscillations · Energy source of limited power · Delay · Elasticity · Damping · Method · Direct linearization

1 Introduction At the present time, humanity is facing acute problems of ecology closely related to energy consumption. The geosphere, atmosphere, aqua sphere, biosphere of the Earth is saturated with a huge amount of industrial, household, chemical waste. Any multielement formations of matter, including the ecology of the Earth, have a limit, upon reaching which irreversible changes occur [1], and this must not be forgotten. In this regard, the world-famous theory of oscillatory systems with limited excitation, set forth by V.O. Kononenko in monographs [2, 3], which systematically substantiated the Sommerfeld effect discovered in experiments in 1902, acquires special significance. There are many works in this direction of the theory of oscillations, in particular [4]. Delay is widespread in technical devices and various technological processes. “Pure lag links are often found in various manufacturing processes where material is moved from one point to another using conveyor belts; in systems for controlling sheet thickness during rolling; in systems of magnetic recording and reproduction, etc.” [5]. Delay in mechanical systems is caused by imperfection of elastic properties of materials, internal friction in them, etc. [6]. The presence of a delay leads to the occurrence of oscillatory © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 187–195, 2021. https://doi.org/10.1007/978-3-030-80472-5_16

188

A. A. Alifov

processes, which can be both useful and harmful. Oscillations in various systems with delay without taking into account the interaction of the oscillatory system and the energy source have been studied in many works, for example, [7–10, etc.]. At the same time, there are only a small number of works devoted to the study of oscillations in systems with delay taking into account this interaction. It is known [11–15] that large labor costs for the analysis of oscillator networks, which play an important role in physics, electronics, neural networks, chemistry, etc., are one of the main problems of nonlinear dynamics. The study of nonlinear dynamic systems is carried out by means of various approximate methods of nonlinear mechanics characterized by laboriousness [16–22]. Methods of direct linearization, which are several orders of magnitude less labor-intensive, differ significantly from them in their ease of use, especially in practice − when calculating technical systems at the design stage [23–27]. The aim of this work is to study forced oscillations under the combined influence of various delays and a limited power energy source based on direct linearization methods.

2 Equations of Motion of the System Consider the dynamic system shown in Fig. 1. The equations of motion of the system have the form m¨x + k x˙ + cx = c1 r1 sin ϕ − f (x)

(1)

I ϕ¨ = M (ϕ) ˙ + c1 r1 (x − r1 sin ϕ)cos ϕ where k is the coefficient of the force of resistance to the movement of the spring, c and c1 are the coefficients of the stiffness of the springs, f (x) is the nonlinear component of the elastic force of the spring with stiffness c, I is the moment of inertia of the motor ˙ is the driving moment of the engine taking rotor rotating the crank with radius r1 , M (ϕ) into account the resistance forces rotor rotation, ϕ˙ - rotation speed.

Fig. 1. System model.

Let’s add (1) terms that take into account delays in the elastic and resistance forces. System (1) under these assumptions will take the form m0 x¨ + k x˙ + c x = λ sin ϕ − f (x) − k x˙  − cτ xτ

(2)

Forced Oscillations with Delays

189

I ϕ¨ = M (ϕ) ˙ + λ (x − r1 sin ϕ) cos ϕ where k and cτ are constants, λ = c1 r1 , x˙  = x˙ (t − ), xτ = x(t − τ ), τ = const and  = const are delays. We represent the nonlinearity f (x) as a polynomial  f (x) = γs x s (3) s

where γs are constants, s = 2, 3, 4,… We now replace the nonlinear function (3) by the linear one according to the method of direct linearization [23], representing in the form F∗ (x) = BF + cF x

(4)

where BF = BF (a), cF = cF (a) are the linearization coefficients determined by the expressions.  Ns γs as , s = 2, 4, 6, . . . (s is even number) (5) BF (a) = s

cF (a) =



N¯ s γs as−1 ,

s = 3, 5, 7, . . . (s is odd number)

s

  Here a = max |x|, Ns = (2r + 1) (2r + 1 + s), N¯ s = (2r + 3) (2r + 2 + s), r is the linearization accuracy parameter, which can be selected from the interval (0, 2), although it has no restrictions [23]. Taking into account (4), Eqs. (2) can be represented in the form x¨ + ω2 x = m−1 L(x, x˙ , ϕ)

(6)

ϕ¨ = I −1 [M (ϕ) ˙ + G(x, ϕ)] where ω2 = c/ m, L(x, x˙ , ϕ) = −k x˙ + λ sin ϕ − BF − cF x − k x˙  − cτ , xτ , G(x, ϕ) = λ(x − r1 sin ϕ) cos ϕ.

3 Method for Solving Equations To solve the first Eq. (6), we can use the method of replacing variables with averaging [23], which allows us to consider stationary and non-stationary processes. Considering an equation of general form x¨ +k x˙ +ω2 x = H (x, x˙ ) with linearized functions, equations of the standard form (ESF) are obtained for determining the nonstationary values of υ and ξ. In this case, the change of variables x = υp−1 cos ψ, x˙ = −υ sin ψ, ψ = p t + ξ,

190

A. A. Alifov

υ = max |˙x| is used. The oscillations amplitude is determined by the expression a = υp−1 . Depending on the H (t, x) function, the p symbol can reflect different frequencies: free oscillations, external influences, self-oscillations. In the presence of an external influence, using the ESF, it is possible to consider the processes in the resonance region and its immediate vicinity. The ESF for determining the non-stationary values of υ and ξ have the form kυ ω 2 − p2 1 dυ dξ = − − Hs (υ, ξ), = − Hc (υ, ξ ) dt 2 dt 2p υ

(7)

where 1 Hs (υ, ξ ) = 2π

2π H (· · · ) sin ψ d ψ, Hc (υ, ξ ) =

1 2π

0

2π H (· · · ) cos ψ d ψ 0

To construct solutions to the second Eq. (6), we use the averaging procedure described in [27]. According to this procedure, for the equation of the energy source of the form J ϕ¨ = M ( ϕ)−R(ϕ, ˙ ϕ, ˙ x, x˙ , x¨ ) using ϕ˙ = + , ˜ where is the main part of the solution,

˜ is the small oscillatory components not taken into account, the equation is obtained 1 d = [M ( ) − R(υ, )] dt J where R(υ, ) =

1 2π

2π 

(8)

R(ϕ, ϕ, ˙ x, x˙ , x¨ ) d ψ.

0

˙ = 0 the equations of stationary motions follow. From (7) and (8) at υ˙ = 0, ξ˙ = 0, kυ + Hs (υ, ξ) = 0, 2

1 ω 2 − p2 − Hc (υ, ξ ) = 0, M ( ) − R(υ, ) = 0 (9) 2p υ

The system of Eqs. (9) allows us to calculate the amplitude a, phase ξ, and velocity for stationary motions, where R(υ, ) represents the load on the energy source.

4 Results of Solving Equations and Stability Conditions Based on the forms (7) and (8), we can write out the dependencies for υ and ξ in the case of (6). Since L(x, x˙ , ϕ) includes the variable ϕ for the energy source, the expressions Hs (υ, ξ) and Hc (υ, ξ) in the case of (6) will take the form Ls (υ, ξ, ), Lc (υ, ξ, ). Using expressions (7) and (8) in relation to the system (6), substituting the frequency p everywhere instead of the frequency and taking into account xτ = υ −1 cos(ψ − τ), x˙  = υ sin (ψ − ), a = υ −1 we obtain the following equations for non-stationary values a, ξ, : da cτ a a λ = − (k + k cos  ) + sin τ − cos ξ dt 2m 2m 2m ω2 − 2 cF cτ λ dξ = + + cos τ + sin ξ dt 2 2m 2m 2ma

(10)

Forced Oscillations with Delays

191

  1 λa d = M ( ) + cos ξ dt I 2 With a small frequency detuning ω − ϕ˙ ∼ or ω − ∼ , where ε is a sufficiently small value, from (10) we obtain the equations of stationary oscillations in the resonance region. Taking into account ≈ ω we have the following expressions for determining the amplitude, phase of oscillations and speed of the energy source.  (11) a2 (F 2 + W 2 ) = λ2 , tgξ = − W F, M ( ) + S( ) = 0 where F = cτ sin ωτ − ω(k + k cos ω ), W = 2 m ω(ω − ) + cF + cτ cos ω τ, S( ) = 0.5a2 F. The S( ) expression reflects the load on the engine from the oscillatory system. As can be seen from (11), in the absence of the nonlinear part of elasticity (f (x) ≡ 0, cF ≡ 0), the amplitude and, accordingly, the load are determined by the expressions   F 2 + W 2 , S( ) = 0.5λ2 F (F 2 + W 2 ) a=λ and the shape of the load curve S( ) becomes similar to that of the amplitude curve. Equations (10) allow us to deduce the stability conditions of stationary resonant oscillations. For this purpose, by composing the equations in variations for (10) and using the Routh-Hurwitz criteria, we obtain the following stability conditions in the resonant zone: D1 > 0, D3 > 0,

D1 D2 − D3 > 0

(12)

where D1 = − (b11 + b22 + b33 ), D2 = b11 b33 + b11 b22 + b22 b33 − b23 b32 − b12 b21 − b13 b31 D3 = b11 b23 b32 + b12 b21 b33 − b11 b22 b33 − b12 b23 b31 − b13 b21 b32     b11 = Q I , b12 = aF 2I , b13 = a2 W 2 I , b21 = 0, b22 = F 2m ω  b23 = −aW 2 m ω, b31 = − 2 m ω, b32 =  ∂ cF = N¯ s γs (s − 1)as−2 , b33 ∂a s

∂ cF 1 (a + W) 2 m ω a ∂a  d M ( ) = F 2m ω, Q = d

5 Calculations To obtain information about the effect of the delay on the characteristics of stationary modes, calculations were performed with the following parameters: ω = 1 s−1 , m = 1 kg f·s2 ·cm−1 , k = 0.02 kgf·s·cm−1 , k = 0.05 kgf·s·cm−1 , cτ = 0.05 kgf·cm−1 , λ = 0.02 kgf, I = 1 kgf · s · cm2 . The delays were chosen from the interval (0, 2π) as p and pτ. The nonlinear component of the elastic force was chosen  in the form f (x) = γ3 x3 , γ3 = ±0.2 kgf · cm−3 and therefore cF = N¯ 3 γ3 a2 , ∂cF ∂a = 2N¯ 3 γ3 a

192

A. A. Alifov

Fig. 2. Amplitude curves: curve 1 – pτ = π/2, curve 2 – pτ = π, curve 3 – pτ = 3π/2.

  where N¯ 3 = (2r + 3) (2r + 5) and N¯ 3 = 3 4 takes place for the linearization accuracy parameter r = 1.5. Note that the number 3/4 is also obtained if the widely used asymptotic method of averaging nonlinear mechanics of Bogolyubov-Mitropolsky is used to solve

Forced Oscillations with Delays

193

Eq. (1) [17]. Therefore, all the results described below take place in the case of solving Eq. (1) by this asymptotic averaging method. Figure 2 shows some results of calculations of amplitudes and their stability under various combinations of delays. Oscillations are stable if the steepness |Q| of the energy source characteristic is within the shaded sector. Note that curve 1 in Fig. 2a coincides with the curve in the absence of delays (cτ = 0, k = 0) and non-linearity (γ3 = 0). From the comparison of the types of characteristics for γ3 > 0 and γ3 < 0, it follows (for example, from the comparison of Fig. 2d and Fig. 2f for p = π/2): the curves are somewhat similar, i.e. they are tilted to the right for γ3 > 0 and to the left for γ3 < 0; there are minor differences in the amplitudes and locations of the characteristics depending on the pτ delay values. Curves 1 and 2 in Fig. 2b are completely unstable for almost all steepness (calculated 0 ≤ |Q| ≤ 300) characteristics of the energy source and are not implemented. Also unstable is the section CE of curve 2 in Fig. 2c and completely unstable curves 1 in Fig. 2d and Fig. 2e, curve 2 in Fig. 2e and curve 1 in Fig. 2f .

Fig. 3. Load curves: p = 0, γ3 = 0.2.

The types of S( ) load curves on an energy source are similar to the types of amplitude curves. For example, in Fig. 3 shows the load curves at p = 0, γ3 = 0.2, corresponding to the amplitude curves in Fig. 2c. The notation of the curves depending on the delay pτ is the same as in Fig. 2c. For p = 0, γ3 = −0.2 the load curves are similar to the curves in Fig. 3, but they tilt to the left.

6 Conclusions The presence of various combinations of delay in the elastic and damping forces can strongly affect the values of the resonant amplitudes of forced oscillations. In addition, depending on these combinations, there may be some displacement of the amplitude curve in the frequency domain. The stability of stationary oscillations and their realizability in real conditions depends both on the properties of the energy source supporting the oscillations and on the combinations of lagging forces. The direct linearization method, which includes the linearization accuracy parameter, and the method based on it for replacing variables with averaging to solve the linearized equation, make it easy to obtain relations for calculating stationary values of the amplitude.

194

A. A. Alifov

References 1. Alifov, A.A.: The fundamental principle which operates the Universe. Regulyarnaya i Khaoticheskaya Dinamika, Moscow-Izhevsk (2012). ISBN: 978-5-93972-948-2, (in Russian) 2. Kononenko, V.O.: Vibrational Systems with Limited Excitation. Nauka, Moscow, Russia (1964). (in Russian) 3. Kononenko, V.O.: Vibrating Systems with Limited Power-Supply. Iliffe, London (1969) 4. Alifov, A.A., Frolov, K.V.: Interaction of Nonlinear Oscillatory Systems with Energy Sources, p. 327. Hemisphere Pub. Corp. Taylor & Francis Group, New York (1990) 5. Theory of automatic control: Textbook. for universities on spec. “Automation and telemechanics”. Part I. Theory of linear automatic control systems / N.A.Babakov, A.A.Voronov, A.A.Voronova and others; ed. Higher school, Moscow, Russia (1986). (in Russian)). 6. Encyclopedia of mechanical engineering. https://mash-xxl.info/info/174754/ 7. Rubanik, V.P.: Oscillations of Quasilinear Systems with Time Lag. Nauka, Moscow, Russia (1969). (in Russian) 8. Zhirnov, B.M.: On self-oscillations of a mechanical system with two degrees of freedom in the presence of delay. J. Appl. Mech. 9(10), 83–87 (1973). (in Russian) 9. Zhirnov, B.M.: Single-frequency resonant vibrations of a frictional self-oscillating system with a delay under an external perturbation. Appl. Mech. 14(9), 102–109 (1978). (in Russian) 10. Astashev, V.K., Hertz, M.E.: Self-oscillation of a visco-elastic rod with limiters under the action of a lagging force. Mashinovedeniye (5), 3–11 (1973). (in Russian) 11. Gourary, M.M., Rusakov, S.G.: Analysis of oscillator ensemble with dynamic couplings. In: AIMEE 2018. The Second International Conference of Artificial Intelligence, Medical Engineering, Education, pp. 150–160 (2018) 12. Ashwin, P., Coombes, S., Nicks, R.J.: Mathematical frameworks for oscillatory network dynamics in neuroscience. J. Math. Neurosci. 6(2), 1–92 (2016) 13. Bhansali, P., Roychowdhury, J.: Injection locking analysis and simulation of weakly coupled oscillator networks. In: Li, P., et al. (eds.) Simulation and Verification of Electronic and Biological Systems, pp. 71–93. Springer (2011) 14. Acebrón, J.A., et al.: The Kuramoto model: a simple paradigm for synchronization phenomena. Rev. Mod. Phys. 77(1), 137–185 (2005) 15. Ziabari, M.T., Sahab, A.R., Fakhari, S.N.S.: Synchronization new 3D chaotic system using brain emotional learning based intelligent controller. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 7(2), 80–87 (2015). https://doi.org/10.5815/ijitcs.2015.02.10 16. Vibrations in technology: directory. Vol. 2. Oscillations of nonlinear mechanical systems // Ed. I.I.Blekhman. Engineering, Moscow, Russia (1979). (in Russian) 17. Bogolyubov, N.N., Mitropolsky, Y.: Asymptotic Methods in the Theory of Nonlinear Oscillations. Nauka, Moscow, Russia (1974).(in Russian) 18. Butenin, N.V., Neymark, Y., Fufaev, N.A.: Introduction to the Theory of Nonlinear Oscillations. Nauka, Moscow, Russia (1976).(in Russian) 19. Tondl A.: On the interaction between self-exited and parametric vibrations. In: National Research Institute for Machine Design Bechovice. Series: Monographs and Memoranda, no. 25, 127 p. (1978) 20. Hayashi, C.: Nonlinear Oscillations in Physical Systems. Princeton University Press, New Jersey (2014) 21. Moiseev, N.N.: Asymptotic Methods of Nonlinear Mechanics. Nauka, Moscow, Russia (1981). (in Russian) 22. Karabutov, N.: Frameworks in problems of structural identification systems. Int. J. Intell. Syst. Appl. (IJISA) 1, 1–19 (2017). https://doi.org/10.5815/ijisa.2017.01.01

Forced Oscillations with Delays

195

23. Alifov, A.A.: Methods of Direct Linearization for Calculation of Nonlinear Systems RCD, Moscow, Russia (2015). (in Russian), ISBN: 978-5-93972-993-2 24. Alifov, A.A.: Method of the direct linearization of mixed nonlinearities. J. Mach. Manuf. Reliab. 46(2), 128–131 (2017). https://doi.org/10.3103/S1052618817020029 25. Alifov, A.A., Farzaliev, M.G.: About the calculation by the method of linearization of oscillations in a system with time lag and limited power-supply. In: Hu, Z., Petoukhov, S., He, M. (eds.) CSDEIS 2019. AISC, vol. 1127, pp. 404–413. Springer, Cham (2020). https://doi.org/ 10.1007/978-3-030-39216-1_37 26. Alifov, A.A.: About direct linearization methods for nonlinearity. In: Hu, Z., Petoukhov, S., He, M. (eds.) AIMEE 2019. AISC, vol. 1126, pp. 105–114. Springer, Cham (2020). https:// doi.org/10.1007/978-3-030-39162-1_10 27. Alifov A.A.: Calculating mixed forced and self-oscillations for delayed elastic constraint and a limited power energy source. J. Mach. Manuf. Reliab. 49(2), 105–109 (2020). https://doi. org/10.3103/S1052618820020053

Method of Static Optimization of the Process of Granulation of Mineral Fertilizers in the Fluidized Bed Bogdan Korniyenko(B) and Lesya Ladieva National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv 03056, Ukraine [email protected]

Abstract. The analysis of approaches to optimization of fertilizer granulation process in the fluidized bed is carried out. The approach to the solution of the problem of static optimization for the process of granulation of mineral fertilizers in a fluidized bed using a mathematical model of the process in the form of criteria dependencies is proposed. It is formulated the criterion of optimality, the necessary conditions of optimality, and the presence of isolated local extremum. The method of the solution of static optimization problem for finding optimal technological and structural parameters for the process of granulation of mineral fertilizers in the fluidized bed is developed. Keywords: Static optimization · Criterion of optimality · Mineral fertilizers · Fluidized bed · Granulation

1 Introduction The development of approaches of effective counteraction to global food crisis belongs to the priority directions of social development in the XXI century. The significant part of the world soil reserves, on which it is possible to grow agricultural products, are focused on the territory of Ukraine. However, their fertility considerably reduced due to intensive exploitation. In these conditions, compensation of the nutrients loss in the soil by conventional fertilizers is insufficient. To provide the necessary conditions of soil formation, it is necessary to create complex mineral and organic mineral fertilizers, the chemical composition of which is determined by the agricultural-ecological conditions of use. The effectiveness of such fertilizers is based on the mineral components uniformity throughout the volume of the granules sizes from 1.5 to 4.5 mm. To achieve these requirements, it is most appropriate to apply the fluidization technique for dehydration of composite solutions to obtain fertilizers with specified properties [1–5]. The presence of the phase transition leads to significant energy losses, that’s why the determining of interconnection between technological parameters and multifactorial processes of granulation at the time of liquid systems dehydration for the purpose of energy consumption optimization, which is an urgent task. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 196–207, 2021. https://doi.org/10.1007/978-3-030-80472-5_17

Method of Static Optimization of the Process

197

2 Review of Optimization Methods of Processes of Dehydration and Granulation in the Fluidised Bed Production of mineral fertilizers at the present stage is characterized by increasing complexity, the presence of a large number of input and output process parameters. Improving the quality of mineral fertilizers, reducing the energy intensity of production is possible when conducting technological processes in certain limited modes. One of the most promising ways to improve the efficiency of devices and related chemical-technological systems is to optimize their operation. Hence the importance of using mathematical modeling methods to describe the processes of production of mineral fertilizers and identify the best conditions for their implementation by solving optimization problems. Granulation processes are the final stage in a number of industries of chemical, microbiological, petrochemical, pharmaceutical, metallurgical, food industries. Widespread use of granulation processes in the production of mineral fertilizers is due to the fact that modern methods of granulation provide a finished product with specified quality indicators. The granular product has good flowability, does not agglomerate, does not dust during transportation and use. One of their most efficient granulation processes is fluidized bed granulation due to high intensity, wide possibilities of automation and process optimization. Development of mathematical models of fluidized bed granulators used to optimize technological and design parameters taking into account technological requirements for the allowable temperature range and moisture content of granules, dispersion of granule sizes, chemical composition of the product, allows to improve the quality of the finished product, reduce energy consumption ultimately gives a great economic effect on the scale of a chemical enterprise. For the production of mineral fertilizers in fluidized bed granulators it is important to determine the optimal feed rate for devices of different heights using mathematical models, which eliminates significant material and time costs for experimental research to select the most favorable modes of operation. Optimization of granulation processes is especially important in connection with the use of automated process control systems that allow you to create embedded optimization systems for different types of granulators. The study of the process of dehydration and granulation in the fluidized bed in terms of optimization requires consideration of three basic elements of the problem: 1) description of possible alternative solutions; 2) definition of the objective function; 3) building a system of restrictions imposed on possible solutions. In the technology of production of mineral fertilizers, one of the main stages of product quality formation is the process of granulation formation with subsequent or simultaneous stabilization of the structure (drying or cooling) and separation of the commodity fraction [4]. It is no coincidence that different schemes of fertilizer production are named after the type of granulator that forms the structure of the production line. In most cases, the methods of granulation have so far been identified and progress in this area is towards modernizing existing equipment in the direction of creating more

198

B. Korniyenko and L. Ladieva

reliable, fairly easy to manufacture and operate structures. Improving the equipment in relation to specific operating conditions has a decisive impact on the efficiency of the production line. However, the modernization of structures should not be one-sided and aimed only at intensifying this process. Ways to change and optimize technology should also be sought to increase process efficiency [8]. The use of fluidization technique to obtain solid composites with specified properties in the presence of phase transitions allows to combine a number of technological stages at a thermal coefficient of more than 60%. Therefore, the creation of mathematical models to create modern process control systems in dispersed systems is relevant. In recent decades, a large number of mathematical models of transfer processes in dispersed systems with different levels of detail have been developed [9]. A static optimization method and a mathematical model of a fluidized bed granulator are used in intensive heat and mass transfer processes in the production of mineral fertilizers for optimization problems. There are a number of approaches to mathematical modeling of dehydration and granulation processes in the fluidized bed. The granular material in the fluidized bed is a chaotic system. Experimental researches have confirmed that local sites of pressure of cavities and concentrations have the disordered fluctuations connected with nonlinear dynamics. Therefore, with the help of chaos, we can describe the dynamics of the fluidized bed and study the processes in the apparatus for different hydrodynamic regimes. Deterministic chaos can occur in a fluidized bed as a result of nonlinear interaction of gas bubbles with granular material. Chaotic hydrodynamics is also used for mathematical modeling of fluidized bed devices [16]. Stochastic mathematical models are applied to systems with random processes, or to complex systems where randomness is used for the necessary simplification. When modeling a fluidized bed, the stochastic approach allows fluctuations in local hydrodynamics or interfacial exchange. When controlling fluidized bed devices, most of the measured parameters have random fluctuations with a fairly high amplitude. Attempts to study the hydrodynamics of multiphase processes in fluidized bed devices using microbalance models are quite effective. These mathematical models solve the energy conservation equation taking into account the interphase interaction. A twophase Euler-Euler model is used to model multifactorial processes of dehydration and granulation in a fluidized bed [19]. For each phase, the mass transfer between the phases, the presence of lifting force and forces acting on the particle - friction forces, pressure forces, gravity, Archimedes forces, adhesion of particles at the phase boundary - were taken into account. The transport equation of the temperature of the granules took into account convective heat transfer, the voltage of the solid phase, the flux of energy fluctuations, the dissipation of the energy of collisions, energy exchange between the phases. This allowed to determine the intensity of the interaction of gaseous (solid) medium and solid particles (dispersed phase) at different hydrodynamic regimes and the corresponding change in the temperature of the granules in the process of dehydration and granulation [13]. The mathematical model [13] comprehensively takes into account the peculiarities of the process, but the large amount of calculation time complicates its use in controlling the process of dehydration and granulation in a fluidized bed in real conditions.

Method of Static Optimization of the Process

199

Therefore, it is proposed to use mathematical models in the form of criterion dependencies.

3 The Optimization of Unit for Dehydration and Mineral Fertilizers Granulation in the Fluidized Bed The article analyzes approaches for optimization of the dehydration and granulation process of mineral fertilizers in the fluidized bed. Also, the optimality criterion is formulated, necessary optimality conditions are obtained, the presence of isolated local extremum is proved and algorithm for solving the static optimization problem for optimal technological and structural parameters for the dehydration processes and granulation of mineral fertilizers in the fluidized bed is developed. We consider the quotient-target analysis for the purpose of energy conservation increasing of the process, which can be used for the control scenario and for building or rebuilding of the unit. For the last one, situational analysis can be used effectively along with it. In terms of energy saving of dehydration and granulation processes in the fluidized bed, it was analyzed possible targets, which can be structured and composed of subtargets. Figure 1 shows the quotient-target diagram for increasing the energy efficiency of the process. The main factors that are needed to be taken into account in the analysis of technicaleconomic efficiency of the process control are: – the energy costs; – the expected duration of the unit work; – the operational costs. The operational costs are often referred to a constant component of the dehydration and granulation processes the cost of goods sold. While control the process, there was not taken into account the capital cost of building the system, the cost of gas distribution grates replacement, cyclone. There are dedicated two goals, each of which consists of sub-goals: C11, C12, C13, C14; C21, C22. The cost of goods sold of the recycling process at the unit can be reduced by increasing the heat generator efficiency, wet cyclone, scrubber, and by the intensification of mass transfer heat in the granulator. Technical-economic indicators of dehydration and granulation in the fluidized bed depend on the kinetics of the process and obtaining the sustainable granulometric compound. Kinetics of dehydration and granulation processes shifts into the dust formation region, if the driving force of the mass transfer process is reduced, which is determined by the difference of the partial pressure of moisture on the surface of the granules and in the heat carrier. This reduces the speed of the moisture removal from the granules. As a result, the mechanical strength of granules is decreasing, which leads to the abrasion and increasing of dust formation [6–18].

200

B. Korniyenko and L. Ladieva

To intensify the heat transfer, it is necessary to increase the thermal pressure. However, the significant increasing of the heat carrier temperature adversely affects the yield of granular product. Because intensifies the process of grinding grains, which would lead to a reduction in the release granular product. It is explained by the formation significant amount of particles smaller than 0.25 mm, which is carried out from the apparatus with the heat carrier.

Fig. 1. The quotient-target diagram of energy efficiency increasing for the dehydration and granulation process of mineral fertilizers in the fluidized bed.

In the layer of large particles d > 1 mm, due to the strong turbulence of the filtering gas flow, convective heat transfers begin to prevail over the thermal conductivity. The higher fluidization speed, the more intensive is the movement of the particles and the higher is the heat transfer coefficient. In terms of energy saving, it is appropriate to consider an optimization of dehydration and mineral fertilizers granulation unit in the fluidized bed for the minimum cost of goods sold of the recycling process. The unit for dehydration and mineral fertilizers granulation in the fluidized bed includes a turbulent burner for natural gas combustion in the surplus of air and obtaining of the heating gases, a gas blower station and an apparatus with the fluidized bed. In order to increase the energy saving level of the mineral fertilizers dehydration and granulation process in the fluidized bed the task of unit work optimization is examined. The specific cost of goods sold of the mineral fertilizers dehydration and granulation process in the fluidized bed is chosen as the optimality criterion, which includes the energy costs of the heating gases pumping and the costs of heating of emulsive phase and solution granules in the fluidized bed. The cost of recycling process in the fluidized bed can be presented by the equation [4]: I = a2 · N2 + a0 · QWG + CPR + C  0 ;

(1)

where a2 - the cost of one kilowatt-hour of electric power that is spent on the gas blower station gear, Euro/(kW·h); N 2 – the electric motor power, that is spent on the gas blower station gear, kW; a0 - the one joule of heat cost that rises with the heating gases, Euro/J; QWG - the amount of warm, that is perceived by an emulsive phase in the granulator, J/h; C PR – the cost of productive charges (that includes depreciation

Method of Static Optimization of the Process

201

deductions, the repair cost and other) Euro/h; C0 - the constant component of the cost of goods sold, Euro/h. As known in the industrial practice, during the thermal apparatus exploitation, productive charges C PR , which include depreciation deductions, often conditionally accept as a constant. But in the apparatus with the surface of heat exchange with the increasing of the heat amount that is passed, depreciation deductions increase. In the apparatus with the fluidized bed, there is no such surface. Therefore, depreciation deductions are considered a constant. Ceteris paribus, the amount of warm, which is perceived by granules from the heating gases, depends on the intensity of heat emission. It can be expressed by correlation: α α QWG = ; or QWG = QWG0 ; QWG0 α0 α0

(2)

where α – the heat emission coefficient from the heat carrier to the granules, at the current value of heat, which is delivered, W/(m2 ·K); α0 – the same as above, but with fixed value, W/(m2 ·K). When (2) is substituted in Eq. (1), the last one looks like: I = a2 · N2 + a0 · QWG0

α + C0. α0

(3)

Dividing the right and the left parts of equation by QWG , we got the goal function, which is formulated as the specific cost of goods sold of the recycling process in the unit for dehydration and mineral fertilizers granulation in the fluidized bed: R = a2

N2 α + a0 · + C0. QWG0 α0

(4)

Supposedly, the warm, which is got from the granules of the heat carrier, is given to the solution on the surface of the granules. Transference of the heat takes place due to the moisture, which evaporates from the granules surface: QWG = W · r;

(5)

where W – the flow of steam from the granules surface, kg/s; r – the hidden warmth from the vaporization, J/kg. Then, with taking (5) into account, Eq. (4) looks like: R = a2

N2 α + a0 · + C0 . W ·r α0

(6)

The above Eq. (6) of expression of the goal function contains the relative change of heat emission αα0 , engine power, which is spent on the gas blower station gear N 2 and the flow of steam from the granules surface W. The value of coefficients of the fixed α0 heat emission and current α can be calculated after the formula [19], represented in the form: α = 1, 6 · 10

−2

 ·

1 ε·a

1/ 3

1 · λb · v / 2 ·

 2/ 3 1 ; d

(7)

202

B. Korniyenko and L. Ladieva

where λb – the thermal conductivity of the fluidized bed environment, W/(m·K); d – the granules diameter, m; a - the thermal diffusivity coefficient, m2 /s; ε – the porosity of the fluidized bed. This formula [19, 20] is valid for an interphase heat exchange in the fluidized bed at Re/ε > 200, where Re – Reynolds number. Let’s mark by index “0” the fixed values of technological and structural parameters, and without an index are their current values. Then: α0 = α



vg0 vg

1/ 3

.

(8)

The flow of steam from the granules surface is described by: W = β · F · p;

(9)

where β – the mass return coefficient, kg/(m2 ·s); F – the mass return surface, m2 ; Δp – the partial pressures difference of moisture vapors on the surface of granules and in the steam-gas mixture, Pa. For determination of the mass return coefficient we used the criterion dependence [19, 20]: 1 Sh = 0, 26 · (Ar · Sc) / 3 ;

(10)

where Sh = Ar = Sc =

β·d D - Sherwood number; g·d 3 ρgr −ρb · ρg - Archimedes νb2 νs D - Schmidt number;

number;

D – the coefficient of granules diffusion in the fluidized bed, m2 /s; νs – the kinematics viscidity of bed, m2 /s, ρ g – the density of heat carrier, kg/m3 , ρ gr – the density of the solid material, kg/m3 , ρ b – the density of emulsive phase, kg/m3 . Wherefrom: 2 β = 0, 26 · D / 3 ·



ρb νb

1/  1 3 ρgr − ρb / 3 · . ρg

(11)

For the estimation of the granules, diffusion coefficient was used the dependence [4]    vg 1 3 D≈ · g·l · −1 ; (12) 60 vkr where l – lesser of the bed sizes (height, width), m; vkr – the speed of the fluidizing beginning, m/s.

Method of Static Optimization of the Process

203

The width of the bed is select as a size of the bed, which equals the apparatus width. The power, which is determined by the gas blower station gear N 2, is determined as follows: N2 = vg · pb · S;

(13)

where Δpb - the pressure difference on the fluidized bed, Pa; S – the area of the gas-distributing grate, m2 . Taking into account (10)–(13) the first member of equation is (6) a2

N2 = β · F · p

1 0, 26 · g / 3 · l ·



vg vkr

vg · pb . 1 2/  1/  3 ρb 3 ρgr −ρb / 3 · · F · (p1 − p2 ) · r −1 · νb ρg (14)

For providing the fluidization, the fluidization number is taken

vg vkr

= 2.

4 The Numerical Analysis Results of the Goal Function During the optimization of the dehydration and granulation processes of mineral fertilizers in the fluidized bed by managing influence, by speed of gas serving, Eq. (6) with taking (8) into account, (14) looks like:   vg0 0,33 −7 vg −6 R = 0, 5 · 10 · + C0 . (15) + 0, 14 · 10 · l vg The presence of the conflict situation in the goal function, which is presented by the Eq. (15), takes place because of it expressed opposite orientation of the specific cost of goods sold components and their non-linearity. In an order to ensure in the presence of an extremum in the considered equation, we conducted a numerical analysis. The results of numerical analysis of the work of unit for dehydration and mineral fertilizers granulation in the fluidized bed are shown on the  Fig. 2. As evidently from the charts of dependence R = f vg have the expressed extreme character. Thus, with the increase of the apparatus width, the points of extremum displaced to the right, in the direction of the higher gas serving speed. The results ground to consider that Eq. (15) can be used for optimization of the dehydration and granulation process of mineral fertilizers in the fluidized bed by the minimum of specific cost of goods sold. It grounds to go to the calculation of optimal values vG and one of basic structural parameters of the apparatus width l. For this purpose, we will use the functions research method of the classic analysis. First derivative of the goal function, which is presented by Eq. (15), being equated to zero, looks like: vg0,33 1 ∂R 0 = 0, 5 · 10−7 · − 0, 0462 · 10−6 · 1,33 = 0. ∂vg l vg

(16)

After the solving this equation relatively to the gas serving speed, it was received: 0,33 0,75 0,33 0,75 vgOPT = 0, 9420,75 · vg0 ·l = 0, 956 · vg0 ·l .

(17)

204

B. Korniyenko and L. Ladieva

Fig. 2. The dependence of the goal function on the gas serving speed.

Accepting vG = 1 m/s and substituting it to (17), it is obtained: vgOPT = 0, 956 · l 0,75 .

(18)

The dependence that we obtained (18), allows calculating vgOPT for different technological and structural parameters. The optimal gas serving speeds, which is received by this equation, comply with the minimum of specific cost of goods sold. In order to ensure in the presence of an optimum, we wrote down the Hesse matrix:

−6 · 1 −0, 5 · 10−7 0, 06 · 10   2,33 vg ; (19) ∇ 2 R vg , l = −0, 5 · 10−7 10−7 · vg which is determined positively. It confirms the presence of the isolated local minimum (15).

5 The Method of Static Optimization for Heat-Mass Exchange Processes The technique of static optimization for heat-mass exchange processes is formulated on the basis of the research: 1. In the target function to take into account the components of cost. 2. Consider operating costs. It is advisable to consider them variables. It makes sense when the heat flow from the coolant significantly affects the wear of heat exchanger surface. Whether it is mixing substances with mixers and transformation of mechanical energy into heat. For substances with great viscosity should be considered variable operating costs.

Method of Static Optimization of the Process

205

3. If the heat transfer occurs not through heat exchanger surface to consider a variable amount of heat that comes from the coolant. 4. Compare the rate of change of the variable heat transfer coefficient to permanent, using the criteria dependencies. If is present stirring compare the rate of change of the engine’s variable power by mixing from constant. 5. For mass exchange processes take into account the variable coefficient of massgiving. 6. It is desirable to achieve the variable technological and constructive parameters in the components of the cost have a degree with different signs. This will ensure the presence of optimum. 7. Check the sufficient conditions of optimality and obtained stationary points for the presence of optimum using the Hesse matrix. Method of solving the problem of static optimization for heat-mass exchange processes using a mathematical model of the process in the form of criteria dependencies can be used at the stage of designing the process, when the variables are technological and constructive parameters [21–28].

6 Conclusion The offered approach to solving the static optimization problem for optimal technological and structural parameters for the dehydration and granulation processes of mineral fertilizers in the fluidized bed by the use of the mathematical model of the process as criterion dependences, that can be used on the stage of process planning, when technological and structural parameters are variables. The local extremum was found analytically, the presence of which was checked by the use of Hesse matrix. The process of dehydration and granulation of mineral fertilizers in a fluidized bed is a complex multifactorial process. Mathematical models of processes for the production of granular fertilizers in a fluidized bed are complex and poorly suitable for use in optimal control systems. Optimization techniques for dewatering and fluidized bed granulation that operate using mathematical partial differential models are only suitable for a narrow range of parameters. The developed method of static optimization allows solving the problem of choosing the optimal parameters of the technological process, which have a major impact on the continuous operation time of the apparatus and the quality of the finished product. Therefore, the use of a mathematical model in the form of criterion dependencies for static optimization is a good solution for systems of optimal control of the production of granular mineral fertilizers in a fluidized bed. The developed method of static optimization can be used in information systems for optimal control of the production of mineral fertilizers in granulators with a fluidized bed.

References 1. Davidson, J., Harrison, D.: Fluidization. Chemistry, Moscow, 725 p. (1973) 2. Grace, J.R.: Fluidized Beds as Chemical Reactors, in Gas Fluidization Technology. John Wiley & Sons, Chichester, New York, Brisbane, Toronto, Singapore, 428 p. (1986)

206

B. Korniyenko and L. Ladieva

3. Kuipers, J.A.M., Hoomans, B.P.B., van Swaaij, W.P.M.: Hydrodynamic models of gasfluidized beds and their role for design and operation of fluidized bed chemical reactors. Fluidization IX. In: Fan, L.-S., Knowlton, T.M. (eds.) Engineering Foundation, New York, pp. 15–30 (1998) 4. Korniyenko, B.: Informacijni tehnologii optymalnogo upravlinnja vyrobnyctvom mineralnyh dobryv. K: Vyd-vo Agrar Media Grup (2014) 5. van Deemter, J.J., Drinkenburg, A.A.H.: In fluidization, pp. 334–347. Netherlands University Press, Amsterdam, Netherlands (1967) 6. Zhulynskyi, A.A., Ladieva, L.R., Korniyenko, B.Y.: Parametric identification of the process of contact membrane distillation. ARPN J. Eng. Appl. Sci. 14(17), 3108–3112 (2019) 7. Galata, L., Korniyenko, B.: Research of the training ground for the protection of critical information resources by iRisk method. In: Zawi´slak, S., Rysi´nski, J. (eds.) Engineer of the XXI Century. MMS, vol. 70, pp. 227–237. Springer, Cham (2020). https://doi.org/10.1007/ 978-3-030-13321-4_21 8. Korniyenko, B.Y., Borzenkova, S.V., Ladieva, L.R.: Research of three-phase mathematical model of dehydration and granulation process in the fluidized bed. ARPN J. Eng. Appl. Sci. 14(12), 2329–2332 (2019) 9. Kornienko, Y.M., Liubeka, A.M., Sachok, R.V., Korniyenko, B.Y.: Modeling of heat exchangement in fluidized bed with mechanical liquid distribution. ARPN J. Eng. Appl. Sci. 14(12), 2203–2210 (2019) 10. Bieliatynskyi, A., Osipa, L., Kornienko, B.: Water-saving processes control of an airport. Paper Presented at the MATEC Web of Conferences, vol. 239 (2018). https://doi.org/10. 1051/matecconf/201823905003 11. Korniyenko, B., Galata, L., Ladieva, L.: Security estimation of the simulation polygon for the protection of critical information resources. Paper Presented at the CEUR Workshop Proceedings, vol. 2318, pp. 176–187 (2018) 12. Galata, L.P., Korniyenko, B.Y., Yudin, A.K.: Research of the simulation polygon for the protection of critical information resources. Paper Presented at the CEUR Workshop Proceedings, vol. 2067, pp. 23-31 (2017) 13. Korniyenko, B., Osipa, L.: Identification of the granulation process in the fluidized bed. ARPN J. Eng. Appl. Sci. 13(14), 4365–4370 (2018) 14. Korniyenko, B., Galata, L.: Implementation of the information resources protection based on the CentOS operating system. In: 2019 IEEE 2nd Ukraine Conference on Electrical and Computer Engineering, UKRCON 2019 - Proceedings 2019, pp. 1007–1011 (2019) 15. Zhuchenko, A.I., Cheropkin, Y.S., Osipa, R.A., Korniyenko, B.Y.: Features of mathematical modeling of the first stage of paper web drying. ARPN J. Eng. Appl. Sci. 15(5), 647–656 (2020) 16. Kornienko, Y.M., Haidai, S.S., Sachok, R.V., Liubeka, A.M., Korniyenko, B.Y.: Increasing of the heat and mass transfer processes efficiency with the application of non-uniform fluidization. ARPN J. Eng. Appl. Sci. 15(7), 890–900 (2020) 17. Korniyenko B., Galata L., Ladieva L.: Research of information protection system of corporate network based on GNS3. In: 2019 IEEE Inernational Conference on Advanced Trends in Information Theory, ATIT 2019 –Proceedings 2019, pp. 244–248 (2019) 18. Korniyenko, B., Galata, L., Ladieva, L.: Mathematical model of threats resistance in the critical information resources protection system. Paper Presented at the CEUR Workshop Proceedings, vol. 2577, pp. 281–291 (2019) 19. Kornienko, Y.M., Sachok, R., Tsepkalo, O.V.: Modelling of multifactor processes while obtaining multilayer humic-mineral solid composites. Chemistry 20(3), E19–E26 (2011) 20. Kornienko, Ya.N., Podmogilnyi, N.V., Silvestrov, A.N., Khotyachuk, R.F.: Current control of product granulometric composition in apparatus with fluidized layer. J. Autom. Inf. Sci. 31(12), 97–106 (1999)

Method of Static Optimization of the Process

207

21. Babak, V., Shchepetov, V., Nedaiborshch, S.: Wear resistance of nanocomposite coatings with dry lubricant under vacuum. Sci. Bull. Natl. Mining Univ. Issue 1, 47–52 (2016) 22. Kravets, P., Shymkovych, V.: Hardware implementation neural network controller on FPGA for stability ball on the platform. In: 2nd International Conference on Computer Science, Engineering and Education Applications, ICCSEEA 2019; Kiev, Ukraine, 26 January 2019– 27 January 2019 (Conference Paper), vol. 938, pp. 247–256 (2019) 23. Karthika, B.S., Deka, P.C.: Modeling of air temperature using ANFIS by wavelet refined parameters. Int. J. Intell. Syst. Appl. (IJISA) 8(1), 25–34 (2016). https://doi.org/10.5815/ ijisa.2016.01.04 24. Ghiasi-Freez, J., Hatampour, A., Parvasi, P.: Application of optimized neural network models for prediction of nuclear magnetic resonance parameters in carbonate reservoir rocks. Int. J. Intell. Syst. Appl. (IJISA) 7(6), 21–32 (2015). https://doi.org/10.5815/ijisa.2015.06.02 25. Malekzadeh, M., Khosravi, A., Noei, A.R., Ghaderi, R.: Application of adaptive neural network observer in chaotic systems. Int. J. Intell. Syst. Appl. (IJISA) 6(2), 37–43 (2014). https:// doi.org/10.5815/ijisa.2014.02.05 26. Bhagawati, K., Bhagawati, R., Jini, D.: Intelligence and its application in agriculture: techniques to deal with variations and uncertainties. Int. J. Intell. Syst. Appl. (IJISA) 8(9), 56–61 (2016). https://doi.org/10.5815/ijisa.2016.09.07 27. Wang, W., Cui, L., Li, Z.: Theoretical design and computational fluid dynamic analysis of projectile intake Int. J. Intell. Syst. Appl. (IJISA) 3(5) 56–632011). https://doi.org/10.5815/ ijisa.2011.05.08 28. Patnaik, P., Das, D.P., Mishra, S.K.: Adaptive inverse model of nonlinear systems Int. J. Intell. Syst. Appl. (IJISA) 7(5), 40–47 (2015)https://doi.org/10.5815/ijisa.2015.05.06

Influence of the Software Development Project Context on the Requirements Elicitation Techniques Selection Denys Gobov1(B) and Inna Huchenko2 1 National Technical University of Ukraine “Igor Sikorsky

Kyiv Polytechnic Institute”, Kyiv, Ukraine [email protected] 2 National Aviation University, Kyiv, Ukraine [email protected]

Abstract. Elicitation (requirements gathering) is a core part of software requirements engineering activities that helps to gather and extract information regarding the current or desired state of enterprise/particular process or product from stakeholders or other sources. Elicited information is used as an input for requirement analysis, design definition, development, and quality assurance activities in IT projects. Different elicitation techniques may be used separately or in conjunction with other techniques to accomplish the elicitation. Best-suited technique choice influences the project management plan, namely its part with the business analysis activities plan and business analysis approach. The goal of the paper is to analyze the influence of project context on requirement elicitation techniques selection, define factors influencing technique selection based on the Machine Learning model, and predict usage of a particular elicitation technique depending on the project attributes and business analyst’s background. Keywords: Software requirements engineering · Machine learning · Decision jungle tree · Business analysis · Document analysis · Interview · Survey · Process analysis · Interface analysis

1 Introduction Requirements engineering is the process of defining, documenting, and maintaining requirements in the engineering design process [1]. The first mentioned activity in this definition is about requirement elicitation. Business analysis is a discipline that extends the scope of requirement engineering activities [2] and contains the following six knowledge areas [3]: Business Analysis Planning, Elicitation, Requirements Life Cycle Management, Strategy Analysis, Requirements Analysis and Design Definition, and Solution Evaluation. Therefore, both disciplines define requirement gathering/elicitation as core activities that are performed to obtain requirements from available sources (stakeholders, document, existing solution, regulation, etc.) and to refine the requirements in detail. The cost of missed or misidentified requirements leads to significant risks. According to [4] © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 208–218, 2021. https://doi.org/10.1007/978-3-030-80472-5_18

Influence of the Software Development Project Context

209

39% of respondents recognized mistakes in the stage of requirements elicitation as one of the most influential factors that led to the failure of software development projects. Other industry studies prove this conclusion as well [5, 6]. Elicitation contains three main tasks: prepare for elicitation, conduct elicitation, and confirm elicitation results. The first task includes elicitation techniques selection. The best practices and recommendations regarding elicitation techniques are defined in the international standard [7] and industrial bodies of knowledge [8–11]. Industrial guidelines and empirical studies define multiple elicitation techniques that have proven themselves in practice. A thorough understanding of the variety of techniques available, their advantages and disadvantages assists the business analyst in adapting to a particular project context [12]. One of the challenges for business analysts/requirement engineers, especially the novice ones, is the selection of the appropriate requirements elicitation techniques that best fit their project. It is difficult to find comprehensive information about when to use a particular technique or a combination of techniques. As a result, some of them are misused, others are never used and only a few are constantly applied. Poorly defined elicitation techniques set leads to gaps in gathered requirements which, in its turn, have a profound negative impact on business analysis and project plan fulfillment, required costs and resources, and project success in general. This study was conducted to analyze the influence of project context attributes on requirement elicitation technique selection in IT projects, built prediction models based on machine learning algorithms, assess the accuracy of these models, and evaluate permutation feature importance. The input data was collected during a survey of practitioners from Ukrainian and international companies with branches in Ukraine involved in the requirement gathering for IT projects [11]. The paper is structured as follows. Section 2 contains a review of the related works on elicitation activities and surveys regarding requirement engineering and business analysis, guidance on their use. Also, in this section, we provide background information on requirements elicitation techniques, collected from industrial bodies of knowledge and study materials prepared by leading international organizations in the business analysis area. Section 3 is devoted to the structure of the questionnaire and survey results. Section 4 concludes the paper with a discussion of the study findings and future work.

2 Related Literature Review There are many studies conducted to analyze the practical and theoretical aspects of elicitation technique selection and to provide some guidance on their use. Dieste and Juristo [13] conducted a systematic review of requirements elicitation techniques based on 26 empirical studies published until the year 2005. They aggregated the results in terms of five guidelines for RE practitioners. In [14] these authors proposed a framework for selecting elicitation techniques based on contextual attributes of the elicitation process and established the adequacy values of each technique for each attribute value. The set of attributes is relevant to the context of the elicitation process and influence the selection of one or other techniques were discovered. Two groups of students were involved in the experiment, practitioners did not take part in the experiment. Ambreen et al. [15] noticed that replicate studies should be conducted in different contexts wherein existing

210

D. Gobov and I. Huchenko

requirement engineers’ interventions were evaluated and implemented in practice. Several studies assess the effectiveness of elicitation techniques in the context of a particular project. Shaheen et al. [16] analyzed the elicitation methods which were used by cloud providers in the Pakistan IT industry. The main result of this work is a statement that one technique can not satisfy project stakeholders. Similar results were received for other techniques (e.g., for crowdsourcing elicitation approach in [17]). Wong and Mauricio [18] defined a set of factors that influenced each activity of the requirements elicitation process and, consequently, the quality: learning capacity, negotiation capacity, permanent staff, perceived utility, confidence, stress, and semi-autonomous. Muqeem and Rizwan [19] proposed a framework that recommends the techniques based on the following components: Pre-Domain Development, Stakeholders Management, Technique Selection, and Prioritization. The main restrictions of the empirical studies mentioned above are the limited number of participants and low practitioners’ involvement. Darwish, Mohamed and Abdelghany [20] proposed a neural network-based model for elicitation techniques selection. The limitation of this model is that the interdependency between elicitation techniques was not analyzed. The following sources were used for creating the elicitation technique list: “A Guide to the Business Analysis Body of Knowledge” (BABOK) from the International Institute of Business Analysis (IIBA), “The PMI Guide to Business Analysis” from the Project Management Institute (PMI), a study guide from the International Requirement Engineering Board (IREB) “Requirements engineering fundamentals” and book “Business Analysis” from British Computer Society (BCS). The analysis of the contents of these sources gives us a set of 16 requirements elicitation techniques: Benchmarking, Brainstorming, Business rules analysis, Collaborative games, Data mining, Design thinking/Lean Startup, Document analysis, Interface analysis, Interview, Observation, Process analysis, Prototyping, Reuse database and guidelines, Stakeholder lists/maps, and personas, Questionnaires or Survey, Workshop.

3 Survey Study A survey description is divided into 3 steps like questionnaire design and data gathering, machine learning model/algorithm selection, and obtained results justification. 3.1 Questionnaire Structure and Content The questionnaire design and data gathering process are described in [21]. The following questions’ categories were included in the questionnaire: • • • • • • •

Q1: General Information. Q2: Requirements Elicitation and Collaboration. Q3: Requirements Analysis and Design. Q4: Requirements Verification and Validation. Q5: Requirements Management. Q6: Attitude to the Business Analysis in the project. Q7: Problems, Causes, and Effects.

Influence of the Software Development Project Context

211

In the given article we focus on the Elicitation and Collaboration topic in the context of general information questions about projects’ backgrounds. Q1: General Information. Questions in this section were intended to give the context such as: • Project size. • The main industrial sector of the current project. The set of industrial sectors was taken from [13] and reworked to domain areas within which services are offered by most of the Ukrainian IT Companies. • Company type: IT or non-IT. For IT companies the separation was made among Outstaff, Outsource, and Product companies. • Company size. • Class of systems/services such as business, embedded, scientific software, etc. • Team distribution (co-located or dispersed). • Primary and secondary roles in the Project. • Experiencein business analyst (BA) or requirements engineer (RE) role. • Certifications. • Way of working in the project (adaptive vs predictive) • Project category for most of the participant’s projects (e.g. greenfield engineering). • BA/RE activities which the respondent is usually involved in. Q2: Requirements Elicitation and Collaboration. Within the given questions category we were interested in elicitation sources, techniques, and project roles that have primary responsibility for the solution requirements (functional, non-functional requirements) elicitation on the respondent’s ongoing project. The following types of elicitation sources were considered: collaborative (relies on stakeholders’ expertise and judgments); experiments, e.g. observational studies, proofs of concept, and prototypes; research, i.e. information from materials or sources that are not directly known by stakeholder. 16 elicitation techniques were proposed as answer options with the ability to select as many as needed for reflecting the full range used by respondents.

3.2 Machine Learning Model Applying Cleaned questionnaire data was processed with machine learning (ML) for getting the recommendations on elicitation techniques usage based on the combinations of factors. MS Azure Machine Learning Studio (classic) [22] was used as the tool for that purpose. Descriptive statistics for the obtained answers are out of the article’s scope. The steps undertaken for the two-classification ML model applying are described below as pseudocode (see Fig. 1).

212

D. Gobov and I. Huchenko

Precondition • Dataset is formed: – Irrelevant records are removed based on the results of the descriptive statistics. – Features F n , n ∈ N (where N is natural numbers set) are selected based on columns corresponding to Q1 and Q2. – Respondents’ answers are transformed into mutually exclusive classes: usage of the particular elicitation technique is set to “1” if the technique is selected, and to “0” if isn’t. • ML algorithm is selected: – The Decision Jungle Tree (DJT) algorithm was empirically selected as the most efficient from the Accuracy and Area Under Curve (AUC) metrics perspective.

In Fig. 2 the ML model steps are shown, where A is the brunch for a full set of factors and B is the brunch for a filtered set of factors based on negative scores analysis. 3.3 Modeling Results Analysis Elicitation techniques were studied with the help of the ML two-class classification model. The following 4 metrics were considered as main for the model accuracy assessment [22]: • Accuracy – a metric that measures the goodness of the classification model as the proportion of true results to total cases. • AUC – a metric that measures area under curve plotted with true positives on the yaxis and false positives on the x-axis. It provides a single number that lets us compare the models of different types. • Precision – a metric that measures the proportion of true results overall positive results. • Recall – a metric measuring the fraction of all correct results returned by the model. Metric’s values for some of the most used techniques according to [21] are listed in Table 1. The accuracy improvement of the selected algorithm is shown in Fig. 3 and Fig. 4 for the Document Analysis elicitation technique. The bold curves on both figures designate the results after features filtering based on negatively scored features exclusion.

Influence of the Software Development Project Context

213

Start (1) (2) (3) (4) (5)

Split data with the coefficient k = 0.75 Train the ML model Test the ML model Access the ML model accuracy considering the Accuracy and AUC metrics Generate and access feature importance scores If iteration i = 1 (6) For every feature Fj, j = 1..n Do check importance score value sj End for; If sj x.reduce (node, target) => (x.getAnyClosest (oldWay).id = = x.nextNodeId)))

AODV Protocol Optimization Software Method

225

This improvement allows in dynamic mode due to the asynchrony of the algorithm without loss of speed, to find the way back to the sender, giving a response to the call to send a new packet, to modify the transmission of messages on the edges of network agents that are trusted and rebuild the current table. The following modification must be added during transmission to eliminate packet looping in the middle of the graph. When sending the message width, index each packet according to the combinations of machine number, node, message number, creation time for the uniqueness of the message. When receiving a message with the top to perform “coloring” of the node relative to the message, marking it as already received, saving the hash of the message body, to eliminate re-looping (Fig. 2). Only hash is stored, takes up little space. Hash comparisons are a quick operation in Redis implementation. If path 1-2-3 is notified faster than a 1–3 packet request, path 3–1 will be blocked.

Fig. 2. Coloring the top of the graph

This improvement along with the hashing of the HPACK algorithm headers, is the most productive due to the avoidance of loops. Header compression also reduces the amount of service traffic, which has a positive effect on the overall trend of the proposed algorithm (Fig. 3). As an improvement, we change the delay time of the message during the processing of the message packet within the node by optimizing the database. The distance-vector algorithm is based on the vector of values of the routing table, which is a two-link list (Fig. 4a). The search is performed by scanning the entire collection of such records in the table, which on a large number of messages and nodes gives complexity n (vector length). Applying a graph-oriented approach in combination with the implementation of Redis hash √ tables and collections and taking it into RAM, we obtain an increase in log(n) − nlog(n) times by optimizing the tables with the scheme shown in Fig. 4, b. By reducing the entities, we reduce the frequency of table updates, respectively, creating opportunities to create an index without losing updates as part of Redis. By providing connected access to each other, we cannot worry about performance. The increment to the logarithmic value is provided by indexing, where the Line table is one key with unique values.

226

L. Oleshchenko and K. Movchan

Fig. 3. Algorithm of the proposed method

a)

b)

Fig. 4. The current scheme of saving the routing table and routing table saving scheme

4 Research Results 4.1 Features of Developed Software Given the possible areas of application of this algorithm, we decided that the system architecture should be built on a client-server model of the modular type. The modular architecture allows implementing the principle of Dependency Injection, which from the point of view of embedding the architecture can in the development of compilation

AODV Protocol Optimization Software Method

227

programming languages with strict typing to provide the resulting software flexibility and support for the implementation of polymorphic hooks. Middleware is internal logic of the application, combined with modules, extensions, databases, internal states of the system (State Machine). Redis is an internal database of the application, which is characterized by extremely high performance in storing and processing data. Document-oriented key storage system is a key-based, open-source dictionary display that can be freely picked up and implemented in the application itself. It is used to optimize headers and as a new data structure instead of a vector representation of standard data structures in the AODV algorithm, to change the two-link list of vertex structures with a full scan. Internal software component RabbitMQ is focused on processing the message queue. Used for packet numbering, queue processing, internal coloring of the graph as part of the interaction-transmission of the message. Inner Queue is a hashed table of message data that stores the general collection of messages that come to the node and are there until the packet is sent. Network Graph Map is a general generated map of the network graph, vertices and edges with their characteristics. Network Part Map is a partial map of the network of adjacent vertices connected to the current vertex, used to optimize intermediate table updates. Connection Manage control is a software application module that is responsible for managing connections to this node. Balancing and route algorithm is implementation of the algorithm implemented in the application. Mobile OS API provides a mobile application for working with embedded systems and the ability to obtain data from them for optimal system operation and a better understanding of the state in which the system itself. • Services is services set and created within the operation of the Android operating system, data interfaces and communication with them. • Sensors is a set of built-in APIs for working with sensors of the device based on Android OS, the set of the latter depends on the device itself. • Wireless is a set of drivers that allows controlling a wireless connection to other devices. • Input is a set of drivers that read all user interactions with the system and the device and provide access to the keyboard and read data from the monitor, microphone, gyroscope, etc. Intertwined with sensors. Hardware API is a set of calls to system and application utilities that allow interacting with the hardware of the system, to create additional daemons inside the Linux machine. This implementation has the following advantages and improvements. The system is not tied to a certain algorithm, the algorithm through the Dependency Injection mechanism can be changed to another, the implementation of a new algorithm is much easier. The logic of working with the system itself is not based on the implementation of an improved algorithm. Detachment allows transferring and implementing the algorithm itself, other than the mobile application, to other systems. The data representation layer is a user interface and is responsible for presenting data to the user and receiving control commands and data from it. The application level implements the basic logic of the application, which carries out the necessary information processing. The data access

228

L. Oleshchenko and K. Movchan

layer provides data storage and access. The developed software application contains a modified algorithm with the ability to interact with the hardware API for Android and supports the exchange of information on MANET network. 4.2 Effectiveness of the Proposed Method To evaluate the method of continuous data transmission, it is necessary to determine the key parameters that are important for the operation of the algorithm in both the minimum mode and the maximum. After a detailed study of the issue, we enter the following indicators: number of nodes involved in routing; each node has a route to any other node in the network; the possibility of loops; the level of knowledge of each node of the topology of the entire network; the number of resources to find a backup communication channel; use of office traffic. We also enter the following quantitative characteristics: number of nodes, time to overcome the path, delay in the node. These characteristics fully assess the improvements in the framework of continuous data transmission in decentralized networks with variable topology. Testing of the proposed method is performed using a computer model that simulated the system. The graph has 12 vertices and approximately the same length of edges. The edges are generated according to the position of the vertices, which, in turn, are placed randomly. Table 1 shows the test results and the results of the comparison of methods. The overall result is the average of all attempts to send a packet from one random node to another in the opposite location with the appropriate initialization. Table 1. Comparison of general characteristics of algorithms Method with modification

Method without modification

Characteristic

7.31

6.15

The number of nodes involved in routing

+

+

Each node has a route to any other node in the network

0.2

0.4

The possibility of loops

+

+

The level of knowledge of each node of the topology of the entire network

352b

438b

Use of office traffic

The results shown in Table 2 can give a clear view that in the range from 40 to 100 kb/s network load, the proposed algorithm also has a number of improvements, namely the reduction of message processing delay in the node measured as the average statistical data testing. This improvement allows to see that delay within the node decreased by 7–15%, depending on the topology of the graph vertex and the interaction of the network agent with other agents; as part of the network operation with the number of nodes that will

AODV Protocol Optimization Software Method

229

Table 2. Time delay of processing in each node Load, kb/s

Modified (delay, s)

Not Modified (delay, s)

20

0.0093

0.0081

40

0.00103

0.00119

50

0.0142

0.0157

70

0.0162

0.0171

100

0.0271

0.0276

120

0.0302

0.0294

increase, this improvement may increase network performance, in particular, its bandwidth; the path time of the message, which depends on the amount of passage and processing of the message by all nodes-transmitters of the network, has also improved (depending on the position of the graph in space and edges). A limitation for the operation of the proposed algorithm is the establishment of a rule for each network node: DB (A) > ... > DB (N ) > ... > DB (B); SB (A) ≤ ... ≤ SB (N ) ≤ ... ≤ SB (B),

(1)

where DB (N ) is the distance to node B in some metric known to node N, and SB (N ) is a measure of the relevance of information about the route to node B, known to node N. The proposed software method differs from existing methods by taking into account changes in the ratio of nodes in the network and adding multiplexing of data transmission between nodes, increases the accuracy when constructing routing tables for data transmission. The future development of the proposed method is carried out by collecting statistics on performance using a data center simulator with current data processing, such as conversion or encoding. A further improvement of proposed method is the implementation of a public key RSA cryptographic algorithm for communication between subscribers based on encryption, testing the breadth of search in an acyclic graph, and the implementation of testing and trust devices in a simulation environment.

5 Conclusions This research presents the traffic routing optimization method based on the AODV algorithm. There are four main indicators of research selected: the speed of establishing the path, the delay of the message in one node, the occurrence of loops, the amount of service traffic. To test the proposed method, a graph simulating a wireless peer-to-peer MANET network of twelve graph vertices is modeled. In the proposed method, each node is assigned a network identifier, which is automatically calculated when connected to it. For each node, the cost and path to the

230

L. Oleshchenko and K. Movchan

neighboring node are calculated, and this data is stored in the routing table. In part, each node can prioritize the most accessible nodes that are closest and have the lowest load and synchronize routing tables. When caching the result, the delay time of the packet before sending to the recipient is reduced. This mechanism allows to quickly restore routes in case of disruption of one of the internal connections. For HTTP protocol stream multiplexing using a binary layer is used, which allows packets to be transmitted in both directions without delays between streams. This has several advantages: parallel multiplexed packet data requests do not block each other; when storing resources on multiple nodes, the speed increases. Compression of data headers by the HPACK reduces service traffic. Due to the use of the created software, the analysis of quality indicators of continuous data transmission and network routing is performed by the proposed method. The proposed method showed the best results by 5–10% reduction of the delay time inside the node and when processing the message, while the amount of service traffic is reduced by 15%. The developed software contains a modified algorithm with the ability to interact with the hardware API for Android and supports the exchange of information on the Ad Hoc Network.

References 1. Khanpara, P., Trivedi, B.: Security in Ad hoc networks. In: Proceedings of International Conference on Communication and Networks, pp. 501–511 (2017) 2. Anchari, A., et al.: Routing problems in mobile Ad hoc Networks (MANET). Int. J. Comput. Sci. Mob. Comput. 6(7), 9–15 (2017) 3. Hara, T.: Data management issues in mobile Ad hoc networks. Proc. Jpn. Acad. Ser. B Phys. Biol. Sci. 93(5), 270–296 (2017). https://doi.org/10.2183/pjab.93.018 4. Farooq, H., Jung, L.T.: Performance analysis of AODV routing protocol for wireless sensor network based smart metering. In: IOP Conference Series: Earth and Environmental Science, 4th International Conference on Energy and Environment, vol. 16 (2013). https://doi.org/10. 1088/1755-1315/16/1/012003 5. Singh, M., Kumar, S.: A survey: Ad-hoc on demand distance vector (AODV) protocol. Int. J. Comput. Appl. 16, 38–44 (2017) 6. Shobha, T., et al.: A reliability based variant of AODV in MANETs: proposal. Anal. Comparison Proc. Comput. Sci. 79, 903–911 (2016). https://doi.org/10.1016/j.procs.2016. 03.112 7. Fang, W., et al.: A source anonymity-based lightweight secure AODV protocol for fog-based MANET. Sensors (Basel) 17(6), 1421 (2017). https://doi.org/10.3390/s17061421 8. Sharma, L., Dimri, P.: An improved AODV with QoS support in mobile Ad-hoc network. In: Proceedings of the 2015 2nd International Conference on Computing for Sustainable Global Development (2015) 9. Oleshchenko, L.: Methods of improving data transmission in dynamic self-organizing networks. In: International scientific and practical conference on Prospects for the development of technical sciences in EU countries and Ukraine» Wloclawek, Republic of Poland, Wloclawek: Izdevnieciba «Baltija Publishing», pp. 16–19 (2018) 10. Sra, P., Chand, S.: QoS in mobile Ad-Hoc networks. Wirel. Pers. Commun. 105(4), 1599–1616 (2019). https://doi.org/10.1007/s11277-019-06162-y

AODV Protocol Optimization Software Method

231

11. Periyasamy, P., Karthikeyan, E.: Energy optimized Ad hoc on-demand multipath routing protocol for mobile ad hoc networks. Int. J. Intell. Syst. Appl. 11, 36–41 (2014). https://doi. org/10.5815/ijisa.2014.11.05 12. Poluboyina, L., Reddy, S.V., Prasad, M.A.: Evaluation of QoS support of AODV and its multicast extension for multimedia over MANETs. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 1, 13–19 (2020). https://doi.org/10.5815/ijcnis.2020.01.02 13. Das, S., Lobiyal, D.K.: Evaluation of message dissemination techniques in vehicular Ad hoc networks using node movement in real map. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 5, 25–33 (2016). https://doi.org/10.5815/ijcnis.2016.05.04 14. Suresh Babu, E., Nagaraju, C., Krishna Prasad, M.H.M.: Empirical performance evaluation of reactive routing protocols for wireless Ad hoc networks in adversarial environment. Int. J. Comput. Netw. Inf. Secur. (IJCNIS), 8(8), 47–58 (2016). https://doi.org/10.5815/ijcnis.2016. 08.06 15. Vaibhav, A., Shukla, D., Das, S., Sahana, S., Johri, P.: Security challenges, authentication, application and trust models for vehicular Ad hoc network– a survey. Int. J. Wirel. Microwave Technol. (IJWMT) 7(3), 36–48 (2017). https://doi.org/10.5815/ijwmt.2017.03.04 16. Chopra, A., Kumar, R.: Efficient resource management for multicast Ad hoc networks: survey. Int. J. Comput. Netw. Inf. Secur. 9, 48–55 (2016). https://doi.org/10.5815/ijcnis.2016.09.07 17. Periyasamy, P., Karthikeyan, E.: Comparative performance analysis of AODV and AODVMIMC routing protocols for mobile Ad hoc networks. Int. J. Comput. Netw. Inf. Secur. 6, 54–60 (2014). https://doi.org/10.5815/ijcnis.2014.06.08 18. Smail, O., Mekkakia, Z., Messabih, B., Mekki, R., Cousin, B.: Energy conservation for Ad hoc on-demand distance vector multipath routing protocol. Int. J. Comput. Netw. Inf. Secur. 6, 1–8 (2014). https://doi.org/10.5815/ijcnis.2014.06.01 19. Madhurya, M., Ananda Krishna, B., Subhashini, T.: Implementation of enhanced security algorithms in mobile Ad hoc networks. Int. J. Comput. Netw. Inf. Secur. 2, 30–37 (2014). https://doi.org/10.5815/ijcnis.2014.02.05 20. Barveen Banu, M., Periyasamy, P.: A survey of unipath routing protocols for mobile Ad hoc networks. Int. J. Inf. Technol. Comput. Sci. 01, 57–67 (2014). https://doi.org/10.5815/ijitcs. 2014.01.07 21. Chitra, M., Siva Sathya S.: Acknowledgement based localization method (ALM) to improve the positioning in vehicular Ad hoc networks. Int. J. Comput. Netw. Inf. Secur. 11, 50–58 (2018). https://doi.org/10.5815/ijcnis.2018.11.06 22. Rakhi, G., Pahuja, L.: Component importance measures based risk and reliability analysis of vehicular Ad hoc networks. Int. J. Comput. Netw. Inf. Secur. 10, 38–45 (2018). https://doi. org/10.5815/ijcnis.2018.10.05 23. Najafi, G., Gudakahriz, S.J.: A stable routing protocol based on DSR protocol for mobile Ad hoc networks. Int. J. Wirel. Microwave Technol. 3, 14–22 (2018). https://doi.org/10.5815/ ijwmt.2018.03.02 24. Ghafouri, A., Ghasemi, A., Ahangar, M.R.H.: A power-based method for improving the ODMRP protocol performance in mobile Ad-hoc networks. Int. J. Wirel. Microwave Technol. 2, 27–36 (2018). https://doi.org/10.5815/ijwmt.2018.02.03 25. Mallapur, S.V., Patil, S.R., Agarkhed, J.V.: A stable backbone-based on demand multipath routing protocol for wireless mobile Ad hoc networks. Int. J. Comput. Netw. Inf. Secur. 3, 41–51 (2016). https://doi.org/10.5815/ijcnis.2016.03.06

Linear Planning Models Using Generalized Production Model of Leontief and Resource Constraints Alexander Pavlov(B)

, Iryna Mukha , Olena Gavrilenko , Liudmyla Rybachuk , and Kateryna Lishchuk

National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv 03056, Ukraine [email protected], [email protected], [email protected], [email protected], [email protected]

Abstract. We consider a problem of forward planning of macroeconomics, both in a deterministic formulation and taking into account the possible uncertainty of the parameters that set the criterion for forward planning’s efficiency. We propose a new method to solve the problem which naturally combines: (a) a modification of the constructive theoretical results obtained earlier in the works of A. Pavlov in the field of optimizing one class of combinatorial problems under uncertainty and (b) possibilities of linear aggregated macroeconomic models of W. Leontief repeatedly confirmed by practice. The results allow: in theoretical terms, to obtain for the first time a constructive condition for the optimal technologies invariance for general resource constraints; to formulate and strictly solve the forward planning problem under uncertainty; to expand significantly the practical application area of the generalized model of W. Leontief. Keywords: Combinatorial optimization · Uncertainty · Macroeconomics · Planning · Aggregation · Production model · Linear programming

1 Introduction The guarantee of stable development of enterprises and economic sectors as a whole significantly depends on the efficiency of forward planning models used [1–11]. Special requirements are imposed on the formalization of problem statements, models, and algorithms that implement forward planning at the macroeconomic level. In particular, an adequate formulation of a prospective macroeconomic development problem should include a possible ambiguity of the parameters of the quality criteria of the models used, as well as an efficient compromise between the adequacy and the simplicity of the used macroeconomic models. The aim of the research in this paper is to significantly expand the area of practical application of the generalized model of W. Leontief using an arbitrary number of resource constraints and to build on its basis a constructive forward planning model for uncertainty conditions. Methodologically, the achievement of this goal is based on systemic combination of theoretical results obtained in [12–15] for © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 232–243, 2021. https://doi.org/10.1007/978-3-030-80472-5_20

Linear Planning Models Using Generalized Production Model

233

the optimization of one class of combinatorial problems under uncertainty and linear macroeconomic models of W. Leontief which have repeatedly proved their efficiency. The results of the paper allow: in theoretical terms, to obtain the first constructive condition for the optimal technologies invariance for general resource constraints; to formulate and strictly solve the forward planning problem under uncertainty; to expand significantly the practical application area of the generalized Leontief’s model.

2 General Theoretical Foundations 2.1 Optimization of Combinatorial Problems Under Uncertainty The following class of combinatorial optimization problems under uncertainty is considered and investigated in [12, 13]: s min wi ki (σ ) (1) σ ∈

i=1

where wi , i = 1, s, are weights, ki (σ ), i = 1, s, is ith arbitrary numerical characteristic of a feasible solution σ from the domain . The coefficients wi , i = 1, s, can take one of L possible sets of values {wil , i = 1, s}, l = 1, L. Further, we employ the following result. Statement 1 [12]. The following is true for arbitrary al > 0, l = 1, L:  s  L s L l arg min = arg min al wil ki (σ ) − fopt al wil ki (σ ) (2) σ ∈

l=1

l = min where fopt

σ ∈

i=1

σ ∈

i=1

l=1

s

l i=1 wi ki (σ ).

2.2 Aggregated Linear Macroeconomic Models of Leontief We will use the following models in this paper: 1. The Leontief open productive model which is specified by a square technological matrix A. The matrix has the largest modulo eigenvalue λ∗ < 1, due to Perron– Frobenius theorem. 2. The Leontief generalized linear production model Aˆ that may contain more than one technology producing one product. The additional property of the model: an arbitrary ˆ = C, yj ≥ 0, j = 1, m (where C admissible basis of a convex compact set Ay is an arbitrary non-zero non-negative vector of products for sale), specifies an open productive model of Leontief.

234

A. Pavlov et al.

3 The Research Methodology The methodology for conducting the theoretical research is as follows: 1. Linear planning models of a quite general form are reduced to a generalized production model of Leontief with additional resource constraints. 2. Application of theoretical results of prof. A. Pavlov [12, 13] to the obtained models allows to obtain: • a constructive generalization of the well-known substitution theorem for the case of two linear resource constraints of a general form—conditions under which arbitrary constraints on the allowable profit do not affect the determination of optimal production technologies (Model 3); • an iterative procedure that takes into account expert constraints imposed on an arbitrary number of resource constraints (Model 2); • a theoretically substantiated efficient analogue of the optimal planning model in the deterministic formulation (Model 1) for the case of forward planning under uncertainty.

4 Forward Planning in a Deterministic Formulation Model 1: min

n i=1

y

pi

m j=1

aˆ ij yj

ˆ ≥ 0, ¯ a0T y ≤ l > 0, yj ≥ 0, j = 1, m Ay n i=1

bit

 m j=1

(3) (4)

 aˆ ij yj  lt > 0, t = 1, M1

(5)

where pi > 0, i = 1, n, is the predicted market value of an ith product unit; bit , l, lt , i = 1, n, t = 1, M1 , are given numbers. Constraints (5) can have, for example, the meaning of predicted logistical constraints that must be satisfied by a vector of products produced for sale C = Aˆ T y. Let yopt be the optimal solution of the linear programming (LP) problem (3)–(5) and ˆ opt . C opt = Ay

(6)

Let us consider the LP problem min y

m j=1

a0j yj

ˆ = C opt , yj ≥ 0, j = 1, m. Ay

(7) (8)

Linear Planning Models Using Generalized Production Model

235

Statement 2. The optimal solution of the LP problem (7), (8) is also the optimal solution of the problem (3)–(5). An arbitrary optimal basis of the LP problem (7), (8) uniquely defines a productive open model of Leontief and an intensity vector y of its technologies work that satisfies the conditions (8). The optimal bases of the LP problem (7), (8), by virtue of the substitution theorem for a generalized arbitrary model, depend only on A, a0 and do not depend on the other numerical parameters of the LP problem (3)–(5). 

Proof. An arbitrary optimal basis of the LP problem (7), (8) uniquely defines a productive open model of Leontief and an intensity vector y satisfying conditions (8), due to the properties of the matrix A. The optimal solution of the LP problem (3)–(5) is obviously a feasible solution of the 

−



LP problem(7), (8) due to the condition (6). The constraints Ay ≥ 0 and constraints (5) are satisfied in the optimal solution of the LP problem (7), (8). The solution corresponds to the optimal value of the functional (3) of the LP problem (3)–(5). When these constraints are satisfied, we have the minimum possible value of the expression a0T y which cannot be greater for any feasible solution of the LP problem (7), (8). Therefore, the optimal solution of LP problem (7), (8) satisfies the constraints a0T y ≤ l > 0. An arbitrary optimal basis of the LP problem (7), (8), due to the well-known substitution theorem, remains optimal for the LP problem m ˆ = C for ∀C ≥ 0. ¯ a0jyj , Ay min j=1

y

Statement 2 is proved. Remark 1. The validity of the substitution theorem follows directly from (a) the wellknown theorem on the optimal and admissible basis of a LP problem given in the standard form and (b) the fact that the inverse matrix of an arbitrary optimal basis of the LP problem (7), (8) is positive. Model 2: yj,

n i=1

pi

m j=1

min

 M2 t=1

j=1,m

wt

 m j=1

 ˆbtj yj − ftopt ,

aˆ ij yj ≥ fpadm > 0, yj ≥ 0, j = 1, m,

m j=1

(9) aˆ ij yj ≥ 0, i = 1, n (10)

n i=1

bti

m j=1

aˆ ij yj  lt , t = 1, M1

(11)

where fpadm > 0 is the given admissible value of the vector of products produced for sale; (11) are the constraints imposed on the vector of products produced for sale; wt >  ˆ 0, t = 1, M2 , are the given expert weights; m j=1 btj yj (btj ≥ 0), t = 1, M2 , are the m resource constraints, one of which is j=1 a0j yj . The resource constraints are measured in the same units. If this is not the case, they are reduced to a common measurement unit 

236

A. Pavlov et al.

(for example, monetary one). This leads to the need of rounding their resulting value to the nearest integer in the corresponding absolute value. fpadm , t = 1, M2 , is the optimal functional value of the following LP problem: m opt ft = min (12) bˆ tj yj , n i=1

pi

yj,

m j=1

j=1

i=1

j=1

aˆ ij yj ≥ fpadm > 0, yj ≥ 0, j = 1, m,

m n

j=1,m

bti

aˆ ij yj ≥ 0, i = 1, n,

m j=1

aˆ ij yj  lt , t = 1, M1 .

Remark 2. An arbitrary optimal solution of the LP problem (9)–(11) satisfies the constraints: m opt opt bˆ tj yj ≥ ft , t = 1, M2 . j=1

Remark 3. It is assumed hereinafter that the LP problem (12) is solvable for t = 1, M2 . We need to solve the following problem: construct an open productive model (modˆ The intensity vector els) of Leontief from the technologies that make up the matrix A. of the model’s (models’) technologies has to be the optimal solution of the LP problem (9)–(11). Also, determine the coefficients of the LP problem (9)–(11), in relation to which the constructed open productive model (models) of Leontief is (are) invariant. By virtue of Statement 1, the LP problem (9)–(11) is equivalent to the following LP problem with the functional  m M2 wt bˆ tj yj (13) min yj,

j=1,m

j=1

t=1

opt

and constraints (10)–(11). Let yj , j = 1, m, be the optimal solution to the LP problem (13), (10), (11). We define the components of the vector C = (C1 , . . . , Cn )T as follows: m opt Ci = aˆ ij yj , i = 1, n. (14) j=1

Let us consider the following LP problem:  m M2 min wt bˆ tj yj , yj,

j=1,m

j=1

t=1

ˆ = C, yj ≥ 0, j = 1, m. Ay

(15) (16)

An optimal solution of the LP problem (13), (10), (11), due to condition (14), is a feasible solution of the LP problem (15), (16). And vice versa: an optimal solution of the LP problem (15), (16), due to condition (16), is a feasible solution of the LP problem (13), (10), (11). Therefore, an optimal solution of the LP problem (15), (16) is at the same time an optimal solution of the LP problem (13), (10), (11).

Linear Planning Models Using Generalized Production Model

237

Statement 3. An arbitrary optimal basis of the LP problem (15), (16), due to the constraints that the matrix Aˆ satisfies, defines an open productive model of Leontief. The vector of its technologies labor intensity is the optimal solution to the LP problem (15), (16) and at the same time the optimal solution to the LP problem (9)–(11). The set of optimal bases of the LP problem (15), (16) and, respectively, the set of open productive models of Leontief are uniquely defined by the matrix Aˆ and the numbers wt bˆ tj , t = 1, M2 , j = 1, m. They do not depend on the values of the other numerical parameters of the LP problem (9)–(11). Proof. The validity of Statement 3 follows from the known theorem on the optimal and admissible basis of an LP problem given in the standard form (see Statement 2). opt

Corollary 1. Let yj , j = 1, m, be an arbitrary optimal solution of the LP problem (13), (10), (11) (yopt in vector form) and t = bˆ Tt yopt − ft

opt

T , bˆ t = bˆ t1 , . . . , bˆ tm , t = 1, M2 .

The following is true if the conditions of corollary 4 of Statement 5 [12] are satisfied. Let there be a solution of the LP problem (13), (10), (11) at fixed values of wt , t = 1, M2 , then



wj − wj j − j < 0, where j is determined by the optimal solution of the LP problem (13), (10), (11) for expert weights wt , t = 1, M2 , t = j, wj = wj . opt

Remark 4. A similar result holds for the LP problem (15), (16) in which ft , t = 1, M2 ,  opt ˆ is a solution to the LP problem ft = min m j=1 btj yj under constraints (16). y

Hence, we have practical recommendations for Statement 3 usage. Suppose we have the desired boundaries lt > 0, t = 1, M2 , on satisfying the constraints bˆ Tt y ≤ lt , t = 1, M2 ,

(17)

and expert priorities for tj , j = 1, M2 , which define the following constraints on expert coefficients wt1 ≥ wt2 ≥ . . . ≥ wtM2 .

(18)

Then, we set the initial set of values wt , t = 1, M2 , at which we solve the LP problem (13), (10), (11) and then solve the LP problem (15), (16) using (14). If the optimal solution of the LP problem (15), (16) violates at least one of the constraints (17), we execute for LP problem (15), (16), without changing constraints (16), the specified number M of iterations (the first of which has already been executed). At each iteration, we increase each of the values of the weights obtained at the previous iteration, for which the corresponding constraint (17) has been violated. As a result, we

238

A. Pavlov et al.

find n technologies, their labor intensity vector that satisfies the conditions (10), (11), and either (a) all the constraints (17) are satisfied, or (b) on a chosen solution from M obtained solutions, each violating at least one of the constraints (17), we achieve

 opt wtj bˆ Ttj ytj ,p − ltj min p=1,M

j

opt opt where ∀tj , bˆ Ttj ytj ,p − ltj > 0, ytj ,p is the optimal solution of the LP problem obtained at the pth iteration.

Remark 5. A theoretically and experimentally efficient procedure of finding arbitrary non-negative expert weights using an empirical matrix of paired comparisons (of arbitrary dimension and possibly not completely filled) is given in [13]. More constructive practical results can be obtained when M2 = 2 (there are two resource constraints on the vector y = (y1 , . . . , ym )T ). Model 3 (two resource constraints). Let’s consider the following LP tasks:

T min w1 bˆ 1 + bˆ 2 y,

(19)



T min w1 bˆ 1 + bˆ 2 y,

(20)

ˆ = C1 , yj ≥ 0, j = 1, n. Ay

(21)



T min bˆ 1 + w2 bˆ 2 y

(22)



T min bˆ 1 + w2 bˆ 2 y,

(23)

ˆ = C2 , yj ≥ 0, j = 1, n Ay

(24)

y

with constraints (10), (11). ˆ opt . Let C1 = Ay 1 y

y

with constraints (10), (11). ˆ opt . Let C2 = Ay 3 y

 ¯ yopt , i = 1, 4, are sets of optimal solutions ¯ bˆ 2 ≥ 0, ¯ b1 , bˆ 2 = 0; where bˆ 1 ≥ 0, i of the LP (19), (10), (11); (20), (21); (22), (10), (11); (23), (24), respectively;  problems opt opt opt yi ∈ yopt i ; f1 , f2 is the optimal functional value of the LP problem (12) for t = 1, 2; {y}f opt , {y}f opt are the sets of their optimal solutions. 

1

2

Statement 4. There are such w1∗ > 0, w2∗ > 0 that for ∀w1 ≥ w1∗ , ∀w2 ≥ w2∗ :

Linear Planning Models Using Generalized Production Model

239

(a) on optimal solutions of the LP problem (19), (10), (11); (20), (21) we have:  opt opt opt ∀y1 ∈ yopt 1 bˆ T1 y1 = f1 (25) opt opt bˆ T2 y1 − f2 = min y∈{y}

opt f1



opt bˆ T y − f , 2

2

 opt opt opt opt opt opt . bˆ T1 y2 = f1 , bˆ T2 y2 = bˆ vT 2 y1 for ∀y2 ∈ y 2 (b) the following is true: opt opt opt ∀y3 ∈ {y}3 bˆ T2 y3 = f2 , opt opt bˆ T1 y3 − f1 = min y∈{y}

opt f2

(26)



opt bˆ T1 y − f1 ,

 opt opt opt opt opt bˆ T2 y4 = f2 , bˆ T1 y4 = bˆ T1 y3 for ∀y4 ∈ yopt 4 . ˆ the open productive models of Leontief corresponding to the (c) For a given matrix A, optimal bases of LP problems (20), (21) or (23), (24) are determined by vectors bˆ 1 , bˆ 2 and do not depend on the values w1 ≥ w1∗ , w2 ≥ w2∗ , as well as on other numerical parameters of the LP problems (19), (10), (11); (20), (21); (22), (10), (11); (23), (24). opt

Proof. Suppose that on an arbitrary optimal solution y1 of the LP problem (19), (10), (11) we have opt opt bˆ T1 y1 = f1 .

(27)

opt

Let y1 ∈ W where W is a finite set of bounded vertices of a convex set defined by constraints (10), (11) which always contains at least one optimal vertex for an arbitrary ¯ functional min C T y, ∀C ≥ 0. y

By virtue the minimum of nonnegative

of Statement

1 [12],  opt opt opt opt opt T T ˆ ˆ + b2 y1 − f2 is reached at y1 under conterms min w1 b1 y1 − f1 y



 opt straints (10), (11). Let y ∈ {y}f opt ∩ W . Then the expression w1 bˆ T1 y − f1 + 1

opt opt bˆ T2 y − f2 = bˆ T2 y − f2 is limited and does not depend on w1 . Therefore, there is such w1∗ > 0 that condition (25) is satisfied for ∀w1 ≥ w1∗ due to the finite number of bounded vertices that form the set W . opt opt The equality bˆ T1 y2 = f1 follows from similar reasoning applied to the LP problem (20), (21). Remark 6. w1∗ is common for the LP problems (19), (10), (11) and (20), (21). This is also true for arbitrary limited admissible (the LP problem (12) is solvable) values of ¯ fpadm > 0, p > 0.

240

A. Pavlov et al.

By virtue of Statement 1 [12] and from the equality of the optimal functional values of the LP problems (19), (10), (11) and (20), (21), follows the validity of remaining equalities of item (a). We can prove the validity of item (b) in a similar way. The validity of item (c) follows from items (a), (b), and the substitution theorem because the optimal functional values of the LP problems (19), (10), (11); (20), (21); (22), (10), (11); (23), (24) remain unchanged for ∀w1 ≥ w1∗ , ∀w2 ≥ w2∗ . Remark 7. Statement 4 can be considered a specific modification of the particular case of the substitution theorem for the case of two resource constraints.

5 Forward Planning Under Conditions of Uncertainty Model 4. The Problem Statement. The uncertainty of Model 1 of forward planning (see Sect. 4) is associated with the component values’ uncertainty in the price vector p > 0¯ of products for sale. Such assumption is natural for the case of forward planning. We formalize the uncertainty of the model by introducing the discrete random variable: ⎛ ⎞ n m   Pi ⎝ aˆ ij yj ⎠ Z = Fopt − i=1

j=1

where (n + 1)-dimensional discrete random variable P1 , . . . , Pn , Fopt is given by the table   l L p1l , . . . , pnl , fopt pl = 1 , l=1 pl > 0, l = 1, L where l fopt = max

n

y

i=1

pil

 m



j=1

aˆ ij yj

(28)

ˆ ≥ 0, ¯ yj ≥ 0, j = 1, m, Ay a0T y ≤ ˆl > 0,

n i=1

bti

 m j=1

 aˆ ij yj  lt > 0, t = 1, M1

(29)

T  where bti , l, lt , i = 1, n, t = 1, M1 are the given numbers, p1l , . . . , pnl is the possible realization of the price vector p of products for sale, l = 1, L. Then the solution of Model 1 of forward planning will have the form: yopt = arg{min MZ} under constraints (29): y

yopt

   L = arg min MZ = arg min y

under constraints (29).

y



l pl fopt l=1

−

n

pl i=1 i

 m j=1

 aˆ ij yj

(30)

Linear Planning Models Using Generalized Production Model

241

By virtue of the corollary to Statement 1 [12], the solution of the combinatorial optimization problem (30) under constraints (29) is obtained as the solution of the following LP problem:   

n L n opt opt T yopt = y1 , · · · , ym = arg max pl pil aˆ ij yj (31) y

i=1

l=1

j=1

under constraints (29). Let ˆ opt . C opt = Ay

(32)

The following obvious analogue of Statement 2 is true. Statement 2a. The optimal solution of the LP problem (31), (29) and, hence, the solution of the combinatorial optimization problem (30), (29) is obtained as the optimal solution of the LP problem (7), (8), for which the vector C is given by expression (32). Its arbitrary optimal basis and the corresponding optimal solution define the productive open model of Leontief and the intensity vector of its technologies. The optimal bases of the LP problem (7), (8), (32) are uniquely defined by the matrix Aˆ and the vector a0 . Remark 8. We can use expert coefficients wl > 0, l = 1, L, in the model (30), instead of the probabilities pl , l = 1, L.

6 Experimental Research All the proposed planning models, except for Model 2, are reduced to solving only the LP problem. Therefore, the efficiency of their implementation corresponds to the efficiency of a general LP problem solving. We give the quantitative characteristics of experimental studies of the iterative procedure (Model 2) in Table 1. Table 1. Experimental studies of the iterative procedure (Model 2) Problem instances number

Products number

Resource constraints number

Maximum iterations (LP problems) number if a solution exists

Average iterations (LP problems) number if a solution exists

100

20–70

5

11

6.13

10

18

9.51

15

27

13.26

242

A. Pavlov et al.

7 Conclusions The scientific novelty and the fundamental difference between the proposed linear models of planning, including forward planning, from all known results (including [1–15]) is in the following. For the first time, the original methodology for studying combinatorial optimization problems under uncertainty developed by prof. A. Pavlov has been applied to linear planning models using the generalized production model of Leontief. This made it possible to obtain three original planning models in deterministic formulation and one efficiently implemented linear model of forward planning under conditions of uncertainty in the values of its linear functional coefficients. For Models 1, 3, and 4, we have determined the invariance conditions for the set of found optimal production technologies. In particular, we have obtained a constructive generalization of the substitution theorem for the case of two resource constraints—conditions under which an arbitrary constraint on the allowable profit does not affect the determination of optimal production technologies. Future research, in addition to theoretical generalizations of the obtained results to the less aggregated linear planning models, suggests the practical implementation of the developed models in real problems of macroeconomic strategic planning.

References 1. Pavlov, A.A., Khalus, E.A., Borysenko, I.V.: Planning automation in discrete systems with a given structure of technological processes. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2018. AISC, vol. 754, pp. 177–185. Springer, Cham (2019). https://doi.org/ 10.1007/978-3-319-91008-6_18 2. Pavlov, A.A.: Long-term operational planning of a small-series production under uncertainty (Theory and Practice). In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2020. AISC, vol. 1247, pp. 167–180. Springer, Cham (2021). https://doi.org/10.1007/978-3-03055506-1_15 3. Zgurovsky, M.Z., Pavlov, A.A.: Algorithms and software of the four-level model of planning and decision making. In: Combinatorial Optimization Problems in Planning and Decision Making: Theory and Applications, Studies in Systems, Decision and Control 1st edn., vol. 173, pp. 407–518. Springer, Cham (2019). https://doi.org/10.1007/978-3-319-98977-8_9 4. Chowdhary, K.R.: Automated planning. In: Fundamentals of Artificial Intelligence. Springer, New Delhi (2020). https://doi.org/10.1007/978-81-322-3972-7_15 5. Ashokkumar, P.R.: Material requirements planning (MRP-I): an overview and methodology for successful implementation. Int. J. Adv. Acad. Stud. 2(2), 24–26 (2020) 6. Titov, V.V., Bezmelnitsyn, D.A., Napreeva, S.K.: Planirovanie funktsionirovaniya predpriyatiya v usloviyakh riska i neopredelennosti vo vneshney i vnutrenney srede. Nauchnotekhnicheskie vedomosti SPbGPU. Ekonomicheskie nauki 10(5), 172–183 (2017). https://doi.org/10. 18721/JE.10516. (in Russian) 7. Jamalnia, A., Yang, J.-B., Feili, A., Xu, D.-L., Jamali, G.: Aggregate production planning under uncertainty: a comprehensive literature survey and future research directions. Int. J. Adv. Manuf. Technol. 102, 159–181 (2019). https://doi.org/10.1007/s00170-018-3151-y 8. Surung, J.S., Bayupati, I.P.A., Putri, G.A.A.: The implementation of ERP in supply chain management on conventional woven fabric business. Int. J. of Inform. Eng. Electr. Bus. (IJIEEB) 12(3), 8–18 (2020). https://doi.org/10.5815/ijieeb.2020.03.02

Linear Planning Models Using Generalized Production Model

243

9. Abusida, A.M., Gültepe, Y.: An association prediction model: GECOL as a case study. Int. J. Inform. Technol. Comput. Sci. (IJITCS) 11(10), 34–39 (2019). https://doi.org/10.5815/iji tcs.2019.10.05 10. Munshi, A., Aljojo, N., Zainol, A., et al.: Employee attendance monitoring system by applying the concept of enterprise resource planning (ERP). Int. J. Educ. Manage. Eng. (IJEME) 9(5), 1–9 (2019). https://doi.org/10.5815/ijeme.2019.05.01 11. Swayamsiddha, S., Parija, S., Sahu, P.K., Singh, S.S.: Optimal reporting cell planning with binary differential evolution algorithm for location management problem. Int. J. Intell. Syst. Appl. (IJISA) 9(4), 23–31 (2017). https://doi.org/10.5815/ijisa.2017.04.03 12. Pavlov, A.A.: Optimization for one class of combinatorial problems under uncertainty. Adapt. Syst. Autom. Control 1(34), 81–89 (2019). https://doi.org/10.20535/1560-8956.1. 2019.178233 13. Pavlov, A.A.: Combinatorial optimization under uncertainty and formal models of expert estimation. Bull. Nat. Tech. Univ. KhPI Ser. Syst. Anal. Control Inf. Technol. 1, 3–7 (2019). https://doi.org/10.20998/2079-0023.2019.01.01 14. Pavlov, A.A., Zhdanova, E.G.: The transportation problem under uncertainty. J. Autom. Inform. Sci. 52(4), 1–13 (2020). https://doi.org/10.1615/JAutomatInfScien.v52.i4.10 15. Pavlov, A.A., Zhdanova, E.G.: Finding a compromise solution to the transportation problem under uncertainty. Adapt. Syst. Autom. Control 1(36), 60–72 (2020). https://doi.org/10.20535/ 1560-8956.36.2020.209764

Programming Style as an Artefact of a Software Artefacts Ecosystem Nikolay Sydorov(B) National Technical University of Ukraine Igor Sikorsky Kyiv Polytechnic Institute, Kyiv 03056, Ukraine

Abstract. In the process of development and maintenance of a software product, many things are created and used that are called software artefacts. Software artefacts are reused, changed, including their relationships, in the process of software product development and maintenance. The complexity and variety of software artefact relationships requires adequate means of their description and management. Software ecosystem research methodology can be used to investigate software artefacts in the context of a software product creation processes. The research subject will be a software artefacts ecosystem. Such an ecosystem is concerned with a more detailed level as compared to a software ecosystem. However, most of the approaches, methods and tools proposed in the software ecosystem research methodology can be also very well used at this level. In the article, the concept of a software artefacts ecosystem is proposed. Concept describes a generic model. According to the software ecosystem research methodology, it is the Cornerstone ecosystem type model, and consists of three actors—platform, software, and artefact. The SD model of the software artefacts ecosystem is described based on the generic model. The roles of the actors in the ecosystem are indicated, the relationships between actors are described. Application of the concept is illustrated by an example of case study of a programming style artefact. When styles (standards) are used, developer’s activities will be more efficient, software is easier to understand, and development and maintenance is cheaper. In a case study, based on the generic model of the software artefacts ecosystem, a declarative model of a programming style ecosystem has been developed. Described a three-level model of a programming style artefact. Tools and processes for creating and using a programming style artefact are developed and described. Keywords: Software engineering · Software artefact · Software ecosystem · Programming · Programming style · Ontology · i*

1 Introduction In the process of development and maintenance of a software product, many things are created and used that are called software artefacts. Artefacts can be different in their form and presentation. They can be part of a software product or provide processes for its development and maintenance. They can represent an intermediate process result or be a part of other artefacts. Thus, there is a huge variety of software artefacts, including © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 244–255, 2021. https://doi.org/10.1007/978-3-030-80472-5_21

Programming Style as an Artefact

245

design plans, work products (specifications, architectural and detailed designs, code, and documentation), user stories, bug reports, including artefacts processing tools, but not limited to this. Various and often complex connections are established between artefacts. In the process of development and maintenance of a product artefacts could be reused and changed, including change of relationships between them. Therefore, artefacts play an important role in the software life cycle regardless of its model and require attention from all interested parties. The complexity and diversity of software artefact relationships requires adequate description and management tools. Software industry is constantly evolving and changing. But not only products and technologies evolve. Many software companies are experimenting with new business models. These experiments lead to a fundamental change in both, company’s and client’s structures. Recently, many companies have been using the concept of a “software ecosystem” to describe their product and own development taking into account customer connections [1]. Ecosystems are recognized as a very promising management tool for an evolving software product [2, 3]. Software ecosystem research methodology can be used to investigate software artefacts in the context of a product creation processes. The research subject will be a software artefacts ecosystem. Such an ecosystem is concerned with a more detailed level compared to a software ecosystem. However, most of the approaches, methods and tools proposed in the software ecosystem research methodology can be very well used at this level. In the article, for the first time, software artefacts ecosystem model is proposed. It defines a goal - that is to study it using case study examples of a programming styles ecosystem based on the author’s works [4–8]. Software artefacts ecosystem generic model is described. In accordance with the software ecosystem research methodology, software artefacts ecosystem is classified as the Cornerstone ecosystem and consists of three actors - platform, software and artefact [9]. The actor’s roles in the ecosystem are indicated, connections be-tween actors are described. Actor types, rules, attributes, their relations and actions can be refined for a specific software artefacts ecosystem model. The same applies to analyzing ecosystem properties. Based on the generic model of the software artefacts ecosystem, a declarative model of the programming style ecosystem has been developed. The programming style is an artefact that plays an important role in the software development and maintenance. Description of the creation and usage processes of the programming style is made using the ontology [10, 11].

2 Related Works 2.1 Software Artefacts In the software life cycle to support the processes of creating and maintaining a software product, many different artefacts are created and used. Wide ranges of components are considered as artefacts, from documentation, work products and their parts, to auxiliary tools. Interacting, artefacts ensure the efficient execution of software life cycle processes.

246

N. Sydorov

In [12], artefacts are analyzed in the context of reuse as equipment in the sense of work [13]. At the same time, three goals (writing, processing and transferring artefacts) and three aspects of equipment (the in-order-to of equipment, readiness-to-hand, presence-and-hand) are considered. In addition, since artefacts are analyzed as reusable components that are embedded in the created software product, their characteristics are taken into account: holism, commonality, reusability and maturity. Considering an artefact as hardware - a thing built into the context of a software product, the interaction of the specified characteristics of software artefacts are investigated. In [14], artefacts are considered in the context of a software product line and are divided into three types - architecture, shared components, and components made from shared ones. For each type of artefacts, three levels of maturity are identified, depending on the degree of integration of the artefact of the corresponding type into the software product line. In [15], artefacts are considered as information parts that are created, modified and used in the RUP processes. Artefacts can be of different types and take different forms, from UML models to executable code, and can be used in the creation and maintenance of a software product. Artefacts are the input and output of actions in RUP processes. In [16], software documentation as an artefact is considered. Artefact as a means of representing information about software is defined. A maintenance model of documentation as a software artefact is introduced. 2.2 Artefacts in Software Development The experience of using artefacts in life cycle processes in several works is considered. In [17], artefact-oriented development of embedded systems is considered. A conceptual model of artefact-oriented development is proposed, examples of its use are given. In [18], artefact-oriented model-driven development is considered. Details a better understanding on how explicating artefacts and their relations facilitates traceability of artefacts, change impact analysis, and interoperability of software tools are considered. The paper [19] concentrates on the paradigm artefact-orientation in requirements engineering and presents a meta model. This meta model is inferred from two concrete domain specific requirements engineering models: one for the application domain of embedded systems and one for the application domain of business information systems. In [20] shown, that ccollaborative development of software products across organizational boundaries in software ecosystems adds new challenges to existing software engineering processes. A new approach offered for handling the diverse software artefacts in ecosystems by adapting features from social network sites. In paper [21], an industrial survey to create an Activity-Based Artefact Quality Model to define what this means from a stakeholder’s viewpoint is proposed. Quality factors of test artefacts that have a positive or negative impact on the activities of Agile testers are explored. Quality model contains 16 quality factors for six test artefacts that are reportedly relevant to at least five stakeholders in the process. In paper [22] reference model and a metamodel for traceability are proposed. The reference model, defined by the conceptual basis, may be used in the creation of traceability approaches. The reference model was used to develop a metamodel. In paper [23], a generic artefact model based on an empirical investigation is proposed. The results of a mapping study in combination with a systematic literature review to analyze the usage of artefacts in agile methods are presented.

Programming Style as an Artefact

247

2.3 Artefact Modeling In the following works, attempts are made to build a model of the artefact. The paper [24] presents a metamodel for software artefacts aiming at providing a new and structured way to represent artefact content, other than current sections hierarchy. This work defines an extension to UML/MOF and SPEM meta-models by means of layers. The paper [25] discusses the theoretical foundations for the representation and interpretation of software artefacts. Based on different levels of perception of artefacts by a person - the user of artefacts introduces three levels of representation of artefacts - physical (physical representation), structural (syntactic structure) and semantic (semantic content). In addition, two steps for processing artefacts - parsing the physical representation, and analyzing the syntactic structure - the result of the first step (interpretation) are introduced. A meta model of artefacts is built on the basis of presentation levels and processing steps. The work [26] considers the architecture of tools that provide the creation and maintenance of metadata about software artefacts, which form an environment consisting of resources - development artefacts. Tools to manage the artefact environment are used. 2.4 Towards a Software Artefacts Ecosystem We are not aware of any work directly devoted to the consideration of problems associated with the study of software artefacts ecosystems. However, there are works, the results of which can be used to solve these problems. In [20], attention is rightly drawn to the fact that in software ecosystems, attention is now paid to the participants only at the top level - these are organizations and teams that create, implement and maintain software products. However, there is a lower level - artefacts, the role of which in the life cycle processes can hardly be overestimated. In [27], there are requirements for describing and analyzing software ecosystems, which in our paper to model software artefacts ecosystems are used.

3 The Generic Model of Software Artefacts Ecosystem This section discusses a generic model of the software artefact ecosystem. Several methods are now used to model software ecosystems [28]. The application of a particular method depends on the type of ecosystem and the goals of the modeling. To represent the software artefacts ecosystem, this work uses the i* modeling approach [28]. In contrast to the most commonly used SSN method, which focuses on describing the software ecosystem at the top level (product, developer, vendor, user), the i* approach provides a description of the ecosystem of a more detailed software presentation layer that corresponds to the level software artefacts. Figure 1 presents generic model of the software artefacts ecosystem. When designing an ecosystem, two groups of requirements are used [27]: descriptive and analytical. The first group includes the requirements for the definition of actors, connections between them and their actions. In addition, the requirements for determining the types, rules and attributes of actors, connections and actions are formulated, as well as the

248

N. Sydorov

Fig. 1. The generic model of the software artefacts ecosystem

requirements for determining the specific characteristics of both the ecosystem as a whole and its elements, for example, productivity, efficiency, security. The second group includes requirements for defining characteristics that provide analysis of the ecosystem from incentives and motivation to sustainability and productivity. Table 1 the actors and roles in the software artefacts ecosystem are given. The software artefacts ecosystem belongs to the Cornerstone type, since the basis of the ecosystem is a technological platform for the development and maintenance of software, the functionality of which is extended by using artefacts [9]. Thus, the actors of the ecosystem are a platform with a management role, software with a software product role, an artefact with a support service provider role. Table 1. Actors and roles in the software artefacts ecosystem Ecosystem type

Actors

The role of the actor in the ecosystem

The Cornerstone ecosystem

Platform

Orchestration

Software

Product

Artefact

Support service provider

Common connections between actors can be indicated (Fig. 2). The platform, in the context of which such components as the life cycle model, organizational and technical support for development and maintenance are considered, defines and uses the artefact as an auxiliary means of implementing processes and filling the structure of a software product. The software depends on the platform, which is the main mean for the implementation of development and maintenance processes. The platform uses the artefact directly as a component in the software structure or indirectly as a means of improving the efficiency of the platform’s processes.

Programming Style as an Artefact

249

The types, rules, and attributes of actors, relationships, and actions can be refined for specific ecosystem models of a software artefact. The same applies to meeting the requirements for the analysis of ecosystem properties [27].

Fig.2. The SD model of the software artefact ecosystem

4 Case Study. The Programming Style Ecosystem Today, methods and tools that are based on reuse have become widespread for the development and maintenance of software products. The application of these methods and tools requires the developer to read, analyze and understand a significant number of representations of work products from different phases of the software life cycle. Reuse is now widespread from requirements specifications to source code and documentation. Therefore, one of the main requirements for software is understandability. The developer’s activities will be more efficient, the software is understandable, and the development and maintenance is cheaper when the styles (standards) are use. They will ensure that the work products of different phases of the life cycle are understandable [25]. Figure 3 shows the model of a programming style ecosystem, which is built based on a generic model of a software artefact ecosystem (Fig. 1). The artefact in this model is the programming style, and the actor, the software, is represented by that part of it - the source code for which the programming style is applied. Artefact - the programming style is platform-specific, as the style rules depend on a number of platform conditions, such as the programming language, management goals, schedule, risks, and project budget. The programming style is used in the source code construction (the phase of software live cycle) and affects the efficiency of the construction and maintenance processes. Based on the artefact model from work [25], described the programming style artefact by the three levels of perception and the two processing steps (Fig. 4).

250

N. Sydorov

Fig. 3. The SD model of programming style ecosystem

Level 1 – Semantic content. The content represents the meaning of an artefact. The content is interpreted in the context of the individual knowledge of the stakeholder (programmer) or the interpreter of the machine (in this case – Protégé). The content is based on the rules of the programming style, which are described by the ontology. Level 2 - Syntactic Structure. The structure of an artefact represents the syntactic expression of its content. The structure of the artefact is described in Web Ontology Language (OWL). Level 3 - Physical Representation. The artefact is represented in the file of OWL text format. There are two the processing steps. Processing Step 1 – Parsing. The outcome of the parsing process is the syntax structure of the artefact. This process is performed by the OWL parser implemented using the OWL API [6]. Processing Step 2 – Interpretation. Interpretation is the process of extracting the content (i.e. the meaning) from the structure. This process is performed by Protégé system. The use of the programming style as an artefact involves the implementation of three processes [29] (Fig. 5): the creation of an artefact, as a result of which the programming language style is built, the use of the style when programs are writing and the process changed of the artefact. In Fig. 6. The ontology of creating a programming style is presented. The ontology describes in detail the participants and actions taking place in this regard in the programming style ecosystem. All ontology concepts are categorized as resources in i* terminology, with the exception of the concept, which represents a goal. At the same time, the concepts Coding phase, Party, Programming language refer to the Platform actor, and the concepts Creating work product style, Style party create guide, Style and Programming language style to the Programming style actor. The ontology of using the programming style is presented (Fig. 7). An ontology describes the relevant actors and activities in a programming style ecosystem. The Party,

Programming Style as an Artefact

251

Protégé Semantics (Content) Interpretation

Meaning

Include

Syntax (Structure) Content structure

Parsing

Composite structure Physical Representation

Fig. 4. The levels of perception of programming style artefact

uses

Acvity occurence

produces

parcipiaon of

Style

Style creaon

Acvity parcipaon

Style change

Style usage

Fig. 5. The processes pattern of style artefact

Coding phase concepts belong to the Platform actor, the Program, Program style concepts to the Software actor, and the Using work product style, Style party using guide, Program language style concepts to the Programming style actor. To implement the processes of creating and using a programming style, tools are created that can be considered, on the one hand, as resources of the Programming style artefact, and on the other hand, as artefacts as part of the Platform artefact. These include the programming style knowledge base and the Reasoner. Thus, the programmer, while coding the program, applies the ontology of the programming style, both for learning the style and for checking the observance of the style in the program. Therefore, two tools are needed - one to create an ontology and support the programmer in the coding process, and the second, to control the application of the programming style in the source code of the program (Fig. 8) [6].

252

N. Sydorov Has- Knowledge-in 1 *

Is- part-of

> Coding phase

*

*

1 > Programming language

> Party > Creating work products style (artefacts)

*

uses

Is- part-of

Programming language style (artefact)

1 Governs

*

1 > Style party create guide

1

1

Is- created- according-to 1 1 > Style

Fig. 6. Ontology of programming style artefact creation Is-part-of

Has- Knowledge-in 1

aquire

1 *

*

*

Party style

Coding phase

> Party

Program

Using work product Style (artefact)

1

1

1

aquire

aquire 1

Governs

* uses

> Style party using guide

Program style

1 Is-

created - according - to

1 Use 1

Programming language style (artefact)

Fig. 7. Ontology of using programming style artefact

The style analyst (may be programmer), using the first tool - Protégé, setting up the ontology to the appropriate programming style, creating a TBox (Fig. 8). After setting up, the programmer is introduced to the programming style with the help of Protégé. The second tool is functionally similar to the reasoner, but adds a function for identifying style errors. In terms of descriptive logic, the reasoner verifies the consistency of the ontology (Fig. 9). Protégé is used to create TBox. It is part of an ontology with terms describing a programming style. Assertions about the source code (ABox) written by the programmer

Programming Style as an Artefact

Programming style (Description)

Protégé (Resource)

Use

Creating

Assist

253

Programming style (Artefact) Level 2 - Syntactic Structure

Ontology Reasoner (Resource) Checking

Task Requirements

Use

Coding

Source Code (Software)

Programmer (Actor)

Fig. 8. Tools usage diagram

Ontology Reasoner (Resource)

Consistency Inconsistency (Style Errors)

Knowledge base of the programming style artefact Level 2 - Syntactic Structure, Level 3 - Physical Representation

TBox

ABox

Terminology axioms

Individual assertions

Language Style Conventions

Source Code (Software)

Fig. 9. Knowledge base of programming style artefact

are created by the corresponding part of the reasoner. It provides the appropriate service using the knowledge base (TBox and ABox). The service includes, firstly, the verification of the consistency of the ontology (a direct function of the reasoner), and secondly, the search for stylistic errors in the source code of the program.

5 Conclusions The research was motivated by an understanding that artefact’s role in the software life cycle is extremely important. Specific attention requires problem of a software artefact

254

N. Sydorov

complexity and diversity. There is also a problem of presenting adequate tools in order to describe and manage artefacts. Existing approaches to the study of artifacts and tools for their use in the context of the software life cycle are aimed at developing software of a certain type, implementing specific processes or using specific development methods (Sect. 2.2). Only a unified, systems approach can provide a fundamental basis for the research and use of artifacts in the software life cycle. The purpose of the research in our article is to show that usage of the software ecosystem research methodology, but at a more detailed software artefact level, is productive. An example of a programming styles ecosystem case study based on the author’s works [4–8]. Using artefact model mentioned in work [25] as a base, programming style artefact is described by the three levels of perception and with two processing steps. The use of the programming style as an artefact is based on the processes pattern [29]. An ontological approach is used to represent the relationship between the programming style and other actors of the ecosystem [6]. Process automation and creation tools are developed and shown to work, as well how to use the artefact in the context of an ecosystem. However, research has shown that actor types, rules and attributes as well as their links and actions in the software attributes ecosystem are different from those in the software ecosystem. Therefore, additional research is needed. In addition, ecosystem’s metric provisioning should be considered in relation to its effectiveness, sustainability and reliability. The use of the of the software artifacts ecosystem as a systems approach to managing artifacts in the context of the software lifecycle will increase the efficiency of its processes.

References 1. Messershmitt, D.G., Messershmitt, C.: Software Ecosystems: Understanding an Indispensable Technology and Industry, 233 p. MIT Press (2003) 2. Lungu, M.F.: Reverse engineering software ecosystems. Doct. Diss. – USI, 208 p. (2009) 3. Guendouz, A., Bennouar, D.: Software product lines composition through partial derivation. IJITCS 9(10), 47–58 (2017) 4. Sydorov, N.: Software Ecology, Software Engineering, pp. 53–61 (2010) 5. Sydorov, N.A.: Software Stylistics Problems of Programming, vol. 2, no. 3, pp. 245–254 (2005) 6. Sidorov, N., Sidorova, N., Pirog, A.: Ontology-driven tool for utilizing programming styles. In: Proceedings of the National Aviation University. N2(71), pp. 84–93 (2017) 7. Sydorov, N., Sydorova, N., Sydorov, E., Cholyshkina, O., Batsurovska, I.: Development of the approach to using a style in software engineering. Eastern-Eur. J. Enterprise Technol. 4/2(100), 41–51 (2019) 8. Sydorov, N.A., Sydorova, N.N., Sydorov, E.N.: Description model of programming style ecosystem. Problems in programming, Special issue. In: Proceeding of the UkrProg’2020, no. 2–3, pp. 74–81 (2020) 9. Knodel, J., Manikas, K.: Towards a typification of software ecosystems. In: Fernandes, J.M., Machado, R.J., Wnuk, K. (eds.) ICSOB 2015. LNBIP, vol. 210, pp. 60–65. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-19593-3_5 10. El mrini, M., El hassan, M., El yassa, M.: A new approach to develop collaborative ontology based BPMN case study: generation of fiscal code. IJITCS 12(1), 1–8 (2020) 11. Sidi, Z., Benslimane, M.: An integrated approach to drive ontological structure from folksonomie. IJITCS 6(12), 35–45 (2014)

Programming Style as an Artefact

255

12. Nuwangi, S.M., Darshana, S.: Software artefacts as equipment: a new conception to software development using reusable software artefacts. In: Thirty-Sixth International Conference on Information Systems, Texas, USA (2015) 13. Heidegger, M.: Being and Time. Translated by John Macquarrie & Edward Robinson. USA: Harper & Row (1927/1962) 14. Bosch, J.: Maturity and evolution in software product lines: approaches, artefacts and organization. In: Chastek, G.J. (ed.) SPLC 2002. LNCS, vol. 2379, pp. 257–271. Springer, Heidelberg (2002). https://doi.org/10.1007/3-540-45652-X_16 15. Rational Unified Process: Best Practices for Software development Teams, Rational Software White Paper TP026B, and Rev. 11/ 01, 18 p. (1998) 16. Glass, R.: Software Maintenance Documentation, SIGDOC ’89, pp. 18–23. Pittsburg, Pennsylvania, USA, ACM Press (1989) 17. Bohm, W., Vogelsang A.: An Artifact-oriented Framework for the Seamless Development of Embedded Systems, Model-Based Engineering of Embedded Systems, pp. 225–234. Springer, Heidelberg (2012) 18. Butting, A., Greifenberg, T., Rumpe, B., Wortmann, A.: On the Need for Artifact Models in Model-Driven Systems Engineering Projects. In: Seidl, M., Zschaler, S. (eds.) STAF 2017. LNCS, vol. 10748, pp. 146–153. Springer, Cham (2018). https://doi.org/10.1007/978-3-31974730-9_12 19. Méndez Fernández, D., Penzenstadler, B., Kuhrmann, M., Broy, M.: A meta model for artefactorientation: fundamentals and lessons learned in requirements engineering. In: Petriu, D.C., Rouquette, N., Haugen, Ø. (eds.) MODELS 2010. LNCS, vol. 6395, pp. 183–197. Springer, Heidelberg (2010). https://doi.org/10.1007/978-3-642-16129-2_14 20. Seichter, D., Dhungana, D., Pleuss, A., Hauptmann, B.: Knowledge management in software ecosystems: software artefacts as first-class citizens. In: ECSA 2010, 23–26 August 2010, Copenhagen, Denmark, pp. 119–126 (2010) 21. Fischbach, J., Mendez, D.: What makes agile test artifacts useful? An activity-based quality model from a practitioners’ perspective. In: ESEM 2020, 8–9 October, Bari, Italy (2020) 22. Azevedo, B., Jino, M.: Modeling Traceability in Software Development: A Metamodel and a Reference Model for Traceability, ENASE, School of Electrical and Computer Engineering, University of Campinas, Brazil, 8p (2015) 23. Kuhrmann, M., Fernández, D.: Towards artifact models as process interfaces in distributed software projects. In: IEEE Workshop Proceedings, p. 10 (2014) 24. Silva, M., Oliveira, T., Bastos, R.: Software artifact metamodel. In: XXIII Brazilian Symposium on Software Engineering, pp. 176–186 (2009) 25. Fernandez, D.M., Bohm, W., Broy, M.: Artefacts in software engineering: a fundamental positioning. Int. J. Softw. Syst. Modeling 26, 9 p. (2018) 26. Dewar, R.G.: Managing Software Engineering Artefact Metadata. Heriot-Watt University, Edinburgh, UK, Department of Computer Science (2005) 27. Sadi, M.H., Yu, E.: Designing software ecosystems: how can modeling techniques help? In: Gaaloul, K., Schmidt, R., Nurcan, S., Guerreiro, S., Ma, Q. (eds.) CAISE 2015. LNBIP, vol. 214, pp. 360–375. Springer, Cham (2015). https://doi.org/10.1007/978-3-319-19237-6_23 28. Yu, E.: Modelling strategic relationships for business process reengineering. Ph.D., thesis. Dept. of Computer Science, University of Toronto (1995) 29. Falbo, R., Nardi, J., Barcellos, M., Guizzardi, G.: Organizing Ontology Design Patterns as Ontology Pattern Languages. Springer, Heidelberg (2011)

Complex Model for Personal Data Management of Online Project Users Solomiia Fedushko1(B) , Oleg Mastykash1 , Yuriy Syerov1 , and Yaryna Kalambet2 1 Lviv Polytechnic National University, Lviv, Ukraine {solomiia.s.fedushko,yurii.o.sierov}@lpnu.ua 2 Ivano-Frankivsk National Technical University of Oil and Gas, Ivano-Frankivsk, Ukraine

Abstract. This article proposes an interdisciplinary solution to the problem of modeling the online project management complex. The implementation of the proposed methods will allow to quickly and efficiently identify and respond to external and internal negative impacts, to develop a plan for overcoming crisis situations, including those caused by the negative psychological, social, and economic impacts of the COVID-19 crisis. The authors analyzed the available software solutions for web content analysis and online project management, formed a list of functional parameters of the online project management complex. Implementation of the developed online project management complex is effective to avoid the emergence of negative audiences to government agencies, reducing the quality of content, spreading fakes, calls to overthrow the constitutional order, separatism or violence, loss of control over online projects due to targeted permanent influence on the audience, Internet bullying, unjustified discrediting, lower reputation, reduction of the flow of income, increase of the cost of maintaining online projects. The obtained results of the analysis of user profile data (comments, shares, likes, and ratings) are the basis for modeling information sharing in web communication services. Keywords: Web project · Management · Online · Project management complex · COVID-19 Crisis · Web communication services

1 Introduction The large amount of data hosted in online project environments, the rapid growth in the number of users, the complexity of building links between users and the extremely wide range of web environments have led to the implementation of the need to design an online project management system. Consolidation of user data and their presentation in the appropriate form for analysis. For web analysis of online project management, the data must be imported into the analytics program in another form. Techniques for visualizing graphs of socially oriented sites have already been developed [1–4], which are formed on the basis of graph modeling theory, data layout methods for creating a two-dimensional graph of a socially oriented site around a selected user of an online project. The described methods and devices are used to identify connections and methods © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 256–269, 2021. https://doi.org/10.1007/978-3-030-80472-5_22

Complex Model for Personal Data Management of Online Project Users

257

of managing online projects or other communication channels. Representing an online project environment in the form of a graph provides an understanding of the scalability of such an environment and the interconnectedness of sites, but this model does not provide comprehensive and specific information about user page blocks and is key in forming links between pages. Two methods of data retrieval were used to study the graph of connections between sites of the online project environment: the method of recursive data retrieval and the method of direct retrieval of data by the input data set. During the implementation of these methods, attention is focused on the details of data retrieval and algorithms for obtaining information from a poorly structured page of the social environment of the Internet. Despite the large number of different solutions for data discovery and analysis in online project environments, the task of building methods of adaptive analysis of inhomogeneous interconnected data in online project environments, in the conditions of constant change of data and their structure, remains unsolved today. All available techniques for analyzing data of online projects only partially solve the problem of collecting information from web environments. Although the technique of collecting information, such as manual page parsing, is universal, it has a significant disadvantage as it is done at extremely low data processing speed and needs a lot of human resources. Existing approaches to automated data collection in online project environments allow you to quickly obtain data, but existing automated collection systems have a number of disadvantages, namely: the specificity and narrowness of the data, convenient data structure for marketing purposes, incomplete picture of online project users, difficulties in adapting methods data management in crisis situations, inaccuracy of data, inability to keep track of the history of data changes of the user of the online project. The primary task of this study is to model the online project management complex as a source of heterogeneous user data. Universal automated data analysis will allow to process large amounts of online project data per time unit. The built models will allow to display a detailed complex of management of web communities and their data. Also, this complex will allow to present heterogeneous data in the form of clearly structured and classified arrays of information, which is one of the necessary conditions for further analysis of these data. The proposed interdisciplinary solution of the problem will allow to quickly and efficiently identify and respond to external and internal negative influences, develop a plan for overcoming crisis situations, eliminating the negative psychological, social and economic consequences of the COVID-19 crisis on online projects.

2 Related Works Scientists Yen H., Hsu S., Huang C. [5] are researching methods of managing online communities to succeed in their management. The methods are based on the theory of behavior of Internet users in society. The methods were tested on 469 respondents and confirm the hypothesis about the importance of norms of cooperation in the community and the level of readiness of technologies for modernization. An important factor in the successful management of online projects is the prediction of crisis situations, which include cyberattacks. A modern approach to analyzing the

258

S. Fedushko et al.

challenges to the security of online projects in cyberspace, which are the main base of cyberattacks. During a nine-month monitoring of fifteen online projects, scientists [6] analyzed 520 messages containing various types of cyberattacks. In the future, scientists plan to analyze their typology, ways to avoid them and methods of reducing the number of cyberattacks. Researchers Yen H., Hsu S., Huang C. [5] are researching methods of managing online communities to succeed in their management. The methods are based on the theory of behavior of Internet users in society. The methods were tested on 469 respondents and confirm the hypothesis about the importance of norms of cooperation in the community and the level of readiness of technologies for modernization. An important factor in the successful management of online projects is the prediction of crisis situations, which include cyberattacks. A modern approach towards analyzing the challenges to the security of online projects in cyberspace are the main reason of cyberattacks. During a nine-month monitoring of fifteen online projects, scientists [6] analyzed 520 messages containing various types of cyberattacks. In the future, scientists plan to analyze the typology of cyberattacks, ways to avoid them and methods to reduce the number of cyberattacks. Sensor web technology has formed the basis for designing a crisis management system for monitoring and protecting the environment [7]. Pearlson Keri E., Saunders Carol S., Galletta Dennis F. [8] were engaged in the development of information systems management strategy. Data verification in e-community management [9–11], analysis of the social communication environment [12–14], and analysis of social media content [15] and search of data in social services [16–18], critical analysis and impact on social users media [19–21], data security and reliability in web communities [22] and virtual project management during the COVID-19 crisis [23–29] are relevant topics for a large number of studies. Despite the availability of a wide range of research [30–35], it is in demand to develop an online project management system, especially in crisis situations.

3 Methods of Study 3.1 Functionality of the Online Project Management Complex There are two key types of users in an online project, namely administrator and user. The IdetityRole entity is responsible for the user type. The user and administrator of the online project platform have access to online project management at different levels. The administrator has access to additional functionality that allows you to use the interface to search for user profiles in online projects. The administrator also has full functionality that allows to manage an online project in real time, which is essential in crisis situations. The online project administrator functionality includes the following functions: – Authorization. – System monitoring. Monitoring includes: getting statistics on program usage, monitoring system loads, and viewing a list of errors. – System settings. Available Settings: Display / hide individual system components, delete users, and Configure System search parameters. – Working with search templates. The administrator can add, edit, and delete templates that are used to search for data by users.

Complex Model for Personal Data Management of Online Project Users

259

– Work with online projects. The administrator can flexibly configure the list of online projects that will be searched for data. This use case is generalized for the following actions: adding a new online project, viewing the list of online projects, and editing online project parameters. – Work with requests from users. The administrator has the ability to interact with users of the system by receiving requests from them and performing a specific response to these requests. User functionality: – Registration. User registration is a two-step process. At the first stage, the user must enter the minimum necessary data to use the system. The next step is to fill in additional information about the online project user. The registration functionality is not available for users who are system administrators. – Authorization. – Data search. This is one of the most difficult parts of the system. The user can search for data through the interface. The search functionality is represented by a multistage process that uses a general algorithm for searching online project data and an algorithm for analyzing photos using machine learning and computer vision. – Work with search results and inquiries. Allows you to consolidate all the data you need to manage online projects. The results are presented as tables and graphs. Allows you to provide communication between system administrators and ordinary users. – Work with other functionality in online project management. Additional functionality that is introduced into the functioning of an online project to increase the efficiency of managing the web community. One of the characteristic features of the online project management complex is that the program provides an opportunity to get the result of analyzing online projects not only in a form that is easy to read to the end user, but also in documents of the following formats:.csv,.xsl,.txt,.json,.xml,.pdf. Only authorized users with administrator rights can use the online project management system. The online project platform has the functionality of registering a new user. All the functionality of online project users is shown in the diagram of variants of the online project management complex in Fig. 1. Decomposition of the functioning processes of the online project management complex is based at dividing the complex into modules: • Client applications: a) A WebApp is a website b) MobileApp is a mobile application c) DesktopApp is a desktop client • WebApi is a web controller module. • Services is a service module. • Modules for working with external SKDs: a) AzureApi is a module for working with the Azure API

260

S. Fedushko et al.

Authorization System settings Add a web project

Specify list of web projects

Edit web project settings

View a list of web projects

Registration

Data search

Monitoring

Work with search templates

Work with web project

Enter search criteria

Download multimedia data

Template selection User



Indication of additional data

Actions with users

Select search type Revision

Delete

Indication of location details

Administrator Working with profiles

Work with search results

Convert Work with requests from users

Comparison of results

Export results

Review

View statistics

Receiving a request

Data change request

Response to the request

Fig. 1. Diagram of options for the online project management complex

b) AwsApi is a module for working with the AWS API c) The Virtual Communities Api is a module for working with Big Data • DAL is an intermediate module for working with data. • Modules for working with CSI: a) Search is a module for searching data b) Parser & Analyzer is a module for data analysis and validation

Complex Model for Personal Data Management of Online Project Users

261

c) Resources is a resource module d) Graphics is a graphics module • Database is a data module. • Background process modules: a) WebService is a background service b) Helpers is an auxiliary module The online project platform is divided into two interfaces: for User work and for Administrator work. That is, the complex is a multi-level system that is built on the principles of Object-Oriented Programming (SOLID). According to the model of the online project Management Complex, 5 layers of the components of the complex are allocated. Layer 1 includes components such as DesktopApp, WebApp, and MobileApp. Layer 2 includes WebApi, Services, SecretsManager, Auth, AzureApi, VirtualCommunityApi, AWSApi, Parser&Analyzer, and Graphics. Layer 3 includes DAL, Database, and Resources. Layer 4 is the layer of background process. This one that includes WebServices and Helpers. Layer 5 is the user authorization management layer. It includes Secrets Management, Signin, and Auth. To work with data, the Dal layer is built, in which data access is implemented through the EntityFramework ORM system. The template - based patterns of Unit OfWork and Repository are chosen as the fundamental patterns of this level. The Facade structural design template is also used in the Software and Hardware Complex, namely, the entry point for all subsystems of the program is the Services Module, which provides access to other modules of the system. The following frameworks and services are also used in the online project management complex: – OAuth. It is used to implement the PAC security. – AWSSDK. It is used to save program resources in the cloud, implement push notifications, and work with databases. – Azure. It is used for modules of computer vision and image analysis. – Mapper. Convert database entities to custom models. – Bootstrapper. – Log4net. Logging program events. – Moq. Simulating program objects that have not yet been created. – NUnitFramework. Writing unit tests for individual program components. – Selenium and PhantomJS. Writing UI tests. The Models module contains unified entities used by the client application layer and the WebAPI module. The main purpose of this module is to present heterogeneous online project data in a structured form and provide interaction between client applications and server components. The WebApi module is an entry point for requests that are executed on behalf of the client applications. This module contains a set of controllers that ensure user interaction with the system. Controllers analyze the request, validate it, and call the necessary part of the system’s functionality.

262

S. Fedushko et al.

Fig. 2. Diagram of the module classes of Models

Complex Model for Personal Data Management of Online Project Users

263

Only this module is available to the client and the rest of the components is not available to him. AdminController is an entity that provides functionality for the administrator’s work. According to the use case diagram, this entity is divided into three partial classes. Each class solves one of the following tasks: working with online projects, working with users, and other methods. 3.2 The Scheme of the Background Process Layer The background process layer is one of the basic layers for consolidating data from heterogeneous online projects. One of the characteristic features of this layer is that its processes will work 24/7, which is key for rapid response to crisis situations in the functioning of an online project. Processes at this level will start during the initial initialization of this system and run continuously until the system is stopped. Although the system already has a profile search toolkit that is available to registered users, this layer is necessary to improve the time indicators of data search. The scheme of operation of profile search processes is shown in Fig. 3. The background process layer is represented by three web services that work independently from each other: – New profiles search service. – Updating existing user data service. – Filtering and syncing existing data service. The processes listed above will not run in parallel, but will run according to a schedule at certain intervals in order to avoid conflicts when working with the resulting data. The search service for new profiles is divided into two processes: – Direct search process for new profiles. – Recursive search. Data for searching for new profiles will be obtained from the table “New user”. When you find information about the user you are looking for, the found tuple entry will be marked as found. The criterion for finding a user is a complete match of the input set of basic attributes. A separate thread is responsible for searching data in a specific online project environment. In each process, threads are generated exactly as many times as there are online projects to analyze. 3.3 User Registration / Authorization Functionality Interface A. User registration takes up to two pages On the first page, the user enters the necessary data for further logging into the system and unique identification. The next page allows the user to enter additional information about themselves, with parameters such as Photo, Last Name, First Name are required, the rest are optional. This step ensures that the background processes of the system are filled with the necessary data for further operation.

264

S. Fedushko et al. The process of direct search Flow 1

Flow 1

Flow N

...

cycle

The process of recursive search

cycle

Flow N

...

cycle

cycle

Obtaining data for the i-th profile from the web project 1

Obtaining data for the i-th profile from the web project N

Obtaining data for the i-th profile from the web project 1

Obtaining data for the i-th profile from the web project N

Data availability check

Data availability check

Data availability check

Data availability check

Reservation

Reservation

Reservation

Reservation

Database

Database

Database

Database

Synchronizing flows

Synchronizing flows

NEW PROFILE SEARCH SERVICE

WEB SERVICES LOAD AND OPERATION MANAGER

DATA SYNCHRONIZATION SERVICE

DATA UPDATE SERVICE

Процес оновлення

Flow 1

Flow 1

Flow N

...

cycle

Процес синхронізації

cycle

cycle

Flow N

... cycle

Search for duplicates of i-th profile from web project 1

Obtaining data for the i-th profile from the web project N

Obtaining data for the i-th profile from the web project 1

Obtaining data for the i-th profile from the web project N

Retrieving data for the i-th profile from web project 1 in the database

Getting data for the i-th profile from web project 1 in the database

Data normalization

Data normalization

Renovation

Reservation

Data comparison

Data comparison

Database

Database

Versioning of updated data

Versioning of updated data

Database

Database

Synchronizing flows

Synchronizing flows

Fig. 3. Scheme of the system operation of background processes

Complex Model for Personal Data Management of Online Project Users

265

The user will not be able to use the system until the second step of registration. In this case, the photo selected by the user will undergo a stage of validation and verification to prove that it is a photo of the actual user. If during the background processes a discrepancy is found in the data entered by the user, then the next logins to the system will be redirected to this page and the user will not be able to use the functionality of the system until he corrects personal data. B. Functionality is available to the authorized user The interface of the User Profiles of the online project page is shown in Fig. 4.

Fig. 4. Top of the Online page of the User Profile

The top of the page includes widgets that contain statistics of the system as a whole. This part of the system is shared by all its users, is filled in automatically and cannot be edited by users. The next part of the page is a list of user profiles of the online project that the user is searching for. The results are data sorted by date of search in descending order.

Fig. 5. List of analyzed users

One of the characteristics of the list of profiles is a link to online project users, in which each existing and found profile is registered. The system also implements the functionality of communication between system users. Communication is implemented

266

S. Fedushko et al.

between system administrators and ordinary users. In turn, ordinary users do not have the ability to communicate with each other. Communication is presented in two formats: – private messages; – system messages. Each message contains the following nodes: – – – – – –

date of departure; status of the message (read / unread); name of the author of the message; image of the author of the message; status of the author of the message (in the system / inactive); message body. It contains text and graphic information.

The system message display area is common to all ordinary users. Only system administrators can fill this area with messages. System messages are read-only to regular users. The Online page of a User displays information about user profiles found in the online project. The data is displayed in the order sorted by search date. The Analysis Results page displays information about the last profile found by the user.

Fig. 6. Page analysis results - related pages

Collecting and processing of consolidated data on online project users makes a significant contribution to simplifying the online project management process. The online project management complex is designed to assist the administrator in managing a virtual project, automating the work for which the contractor devotes a lot of time and money.

Complex Model for Personal Data Management of Online Project Users

267

Fig. 7. Analysis results page - other information

4 Conclusion The online project management complex is useful for such areas of professional activity in which there is a need for research and analysis of personal data of online project users. For example, the emergence of a negative audience for government agencies, reduced quality of content (distribution of fakes), illegal activities (calls to overthrow the constitutional order, separatism or violence), loss of control over online projects due to targeted permanent impact on the audience, which leads to negative consequences for online projects and their users (Internet bullying, unjustified discrediting), to reduce the reputation, reduce the flow of profits, increase the cost of maintaining online projects. The obtained results of the analysis of user profile data (comments, distributions, likes and ratings) are the basis for modeling of the information sharing in web communication services. The methods proposed in this work for online project management are intended for implementation in such areas of human activity as, for example, administration of an educational institution (search of potential entrants, analysis of information portrait of students), marketing activities of the company. external and internal outflows). In future research, it is planned to build information profiles of web project users on the basis of a complex model for personal data management of online project users in order to verify their data. Acknowledgment. This research is supported by National Research Foundation of Ukraine within the project “Methods of managing the web community in terms of psychological, social and economic influences on society during the COVID-19 pandemic”, grant number 94/01–2020.

References 1. Nie, Z.: Automated social networking graph mining and visualization. U.S. Patent No 9,990,429 (2018) 2. Mckee, M.W., Schultz, P.T., Sartini, R.A.: Weighting social network relationships based on communications History. U.S. Patent No 9, 400, 972 (2016) 3. Valente, T.W., et al.: Social Network Analysis for Program Implementation. Plos One 10(6), E0131712 (2015) 4. Digital 2019: Global Internet Use Accelerates [wearesocial]. – Mode of access. https://wea resocial.com/blog/2019/01/digital-2019-global-internet-use-accelerates

268

S. Fedushko et al.

5. Yen, H.R., Hsu, S.H.Y., Huang, C.Y.: Good soldiers on the Web: understanding the drivers of participation in online communities of consumption. Int. J. Electron. Commer. 15(4), 89–120 (2011) 6. Viduka, D., Kraguljac, V., Lavrni´c, I.: Security challenges behind the development and increased use of open source web content management systems. J. Inf. Organ. Sci. 44(1), 185–198 (2020) 7. Markovi´c, N., Stoimenov, L.: Sensor web architecture for crisis management. In: ICEST 2009, 25–27 June, pp. 263–266 (2009) 8. Pearlson, K.E., Saunders, C.S., Galletta, D.F.: Managing and Using Information Systems: A Strategic Approach. Wile, Hoboken (2019) 9. Fedushko, S., Syerov, Yu., Skybinskyi, O., Shakhovska, N., Kunch, Z.: Efficiency of using utility for username verification in online community management. In: Proceedings of the International Workshop on Conflict Management in Global Information Networks (CMiGIN 2019), Lviv, Ukraine, November 29, 2019. CEUR-WS.org, vol. 2588, pp. 265–275 (2020). http://ceur-ws.org/Vol-2588/paper22.pdf 10. Timonin, A.Y., Bozhday, A.S., Bershadsky, A.M.: The process of personal identification and data gathering based on big data technologies for social profiles. In: Chugunov, A.V., Bolgov, R., Kabanov, Y., Kampis, G., Wimmer, M. (eds.) DTGS 2016. CCIS, vol. 674, pp. 576–584. Springer, Cham (2016). https://doi.org/10.1007/978-3-319-49700-6_57 11. Papadopoulos, S.: Overview of the special issue on trust and veracity of information in social media. ACM Trans. Inf. Syst. (TOIS) 34(3), 14 (2016) 12. Lodhia, S., Stone, G.: Integrated reporting in an internet and social media communication environment: conceptual insights. Aust. Account. Rev. 27(1), 17–33 (2017) 13. Serrat, O.: Social network analysis. In: Knowledge solutions, pp. 39–43. Springer, Singapore (2017). https://doi.org/10.1007/978-981-10-0983-9_9 14. Schivinski, B., Dabrowski, D.: The effect of social media communication on consumer perceptions of brands. J. Mark. Commun. 22(2), 189–214 (2016) 15. Valenzuela, S., Correa, T., Gil De Zuniga, H.: Ties, likes, and tweets: using strong and weak ties to explain differences in protest participation across Facebook and Twitter use. Polit. Commun. 35(1), 117–134 (2018) 16. Kanuri, V.K., Chen, Y., Sridhar, S.: Scheduling content on social media: theory, evidence, and application. J. Mark. 82(6), 89–108 (2018) 17. Peleshchyshyn, A.M., Mastykash, O.V.: General algorithm for searching user data in social media of the Internet. Eur. J. Eng. Res. Sci. 5(1), 82–86 (2020) 18. Tang, X.: Chinese website classification method and system based on characteristic analysis of website homepage. U.S. Patent Application No 15/325,083 (2017) 19. Marra, F.: Detection of GAN–generated fake images over social networks. In: 2018 IEEE Conference on Multimedia Information Processing and Retrieval (MIPR), pp. 384–389. IEEE (2018) 20. Sivarajah, U.: Critical analysis of Big Data challenges and analytical methods. J. Bus. Res. 70, 263–286 (2017) 21. Muhammad, S.S., Dey, B.L., Weerakkody, V.: Analysis of factors that influence customers’ willingness to leave big data digital footprints on social media: a systematic review of literature. Inf. Syst. Front. 20(3), 559–576 (2018) 22. Ma, J.: Balancing user profile and social network structure for anchor link inferring across multiple online social networks. IEEE Access 5, 12031–12040 (2017) 23. Zhang, Z., Gupta, B.B.: Social media security and trustworthiness: overview and new direction. Futur. Gener. Comput. Syst. 86, 914–925 (2018) 24. Krausz, M., Westenberg, J.N., Vigo, D., Spence, R.T., Ramsey, D.: Emergency response to COVID-19 in Canada: platform development and implementation for eHealth in crisis management. JMIR Public Health Surveillance 6(2), e18995 (2020)

Complex Model for Personal Data Management of Online Project Users

269

25. Brandusescu, A., Sieber, R.E., Jochems, S.: Confronting the hype: the use of crisis mapping for community development. Convergence 22(6), 616–632 (2016) 26. Hacker, J., vom Brocke, J., Handali, J., Otto, M., Schneider, J.: Virtually in this together– how web-conferencing systems enabled a new virtual togetherness during the COVID-19 crisis. Eur. J. Inf. Syst. 1–22 (2020) 27. Kryvenchuk, Y., Boyko, N., Helzynskyy, I., Helzhynska, T., Danel, R.: Synthesis control system physiological state of a soldier on the battlefield. In: CEUR, vol. 2488. Lviv, Ukraine, pp. 297–306 (2019) 28. Molnár, E., Molnár, R., Kryvinska, N., Greguš, M.: Web Intelligence in practice. The Society of Service Science. J. Serv. Sci. Res. 6(1), 149–172 (2014). ISSN: 2093–0720, Journal no. 12927 (2014) 29. Aneta, P.-M., Kaczmarek, D., Kryvinska, N., Fatos, X.: Studying usability of AI in the IoT systems/paradigm through embedding NN techniques into mobile smart service system. Springer J. 101(11), 1661–1685 (2019). https://doi.org/10.1007/s00607-018-0680-z. ISSN: 0010–485X, Computing, November 30. Baako, I., Umar, S.: An integrated vulnerability assessment of electronic commerce websites. Int. J. Inf. Eng. Electron. Bus. (IJIEEB) 12(5), 24–32 (2020). https://doi.org/10.5815/ijieeb. 2020.05.03 31. Ayeni, O.A., Mercy, A., Thompson, A.F., Mogaji, A.S.: Web-based student opinion mining system using sentiment analysis. Int. J. Inf. Eng. Electron. Bus. (IJIEEB) 12(5), 33–46 (2020). https://doi.org/10.5815/ijieeb.2020.05.04 32. Haroon, M.: Comparative analysis of stemming algorithms for web text mining. Int. J. Modern Educ. Comput. Sci. (IJMECS) 10(9), 20–25 (2018). https://doi.org/10.5815/ijmecs.2018. 09.03 33. Amjad, M., Linda, N.J.: A web based automated tool for course teacher evaluation system (TTE). Int. J. Educ. Manag. Eng. (IJEME) 10(2), 11–19 (2020). https://doi.org/10.5815/ijeme. 2020.02.02 34. Mithun, A.M., Bakar, Z.A.: Empowering information retrieval in semantic web. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 12(2), 41–48 (2020). https://doi.org/10.5815/ijcnis.2020. 02.05 35. Fotsoh, A., Sallaberry, C., Lacayrelle, A.L.P.: Retrieval of complex named entities on the web: proposals for similarity computation. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 11(11), 1–14 (2019). https://doi.org/10.5815/ijitcs.2019.11.01

Application of a Modified Logarithmic Rule for Forecasting Oleksandr Olefir(B) and Orest Lastovsky National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv 03056, Ukraine

Abstract. The paper analyzes the mathematical methods of forecasting in market conditions, explores the principle of operation of prediction market and finds that using the prediction market model as a forecasting tool is appropriate, and that different methods of data aggregation should be used depending on the liquidity and number of active users. When comparing the modified logarithmic market valuation rule with the traditional one, we get an automatic price adjustment by an automated market maker, depending on activity of users and a lower loss of funds in the worst-case scenario. This way, prediction market model will become more practical to use, reducing the dependence of forecast quality on the number of users and mitigating the effects of potential losses. Keywords: Prediction market · Logarithmic rule · Market maker · Forecasting · Regression · Boosting

1 Introduction Predictin markets (PM) are defined as markets of contracts (more precisely, a subset of contracts on market of futures) with trading the outcomes of the uncertain events expected in the future. They are increasingly recognized as an instrument for effective dissemination of information about future events and are likely to become even more important in the modern world of data and statistics [1]. PMs provide information on the costs, benefits, or likelihood of an incident to occur. Participants trade contracts that value future events with predominantly binary results. Historically, typical topics in prediction markets have been, for example, forecasting the likelihood of presidential re-election or predictions about climate change, constitutional referendums, gas prices, disease outbreaks, expected countries’ economic development, or earthquakes. The idea of using prediction markets in management is relatively new, but it has been proven that they can be used for a very dynamic sales forecasting or prediction of any other key performance indicators. A particular advantage of prediction markets is the accuracy of their forecasts compared to the predictive power of experts or surveys, which empirically tend to be less effective [2]. As a consequence, it is no wonder that prediction markets are likely to be used in a wide variety of contexts. Well-known prediction markets are, for example, © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 270–280, 2021. https://doi.org/10.1007/978-3-030-80472-5_23

Application of a Modified Logarithmic Rule for Forecasting

271

electronic markets that deliver forecasts of the US Presidential elections that are more accurate than the results of parallel opinion polls, such as Gallup Poll [3]. Moreover, the accuracy of forecasts, creativity and anonymity, which are the characteristics of prediction markets, motivate the publication of knowledge even by those people who usually refuse to participate in organizational systems of knowledge management. Participation in the forecasting market is a creative experience as people think about future events and potential incidents, and act as traders in dynamic and fluctuating conditions. Electronic platform functions as a virtual marketplace where participants use their money to trade their knowledge and expectations about specific events. The gaming nature of prediction markets can be further illustrated by the use of gambling money or “game currency”, which can then be converted into real prizes. The main purpose of organization of prediction market for organizational purposes is to gain the wisdom of many people. Knowledge disseminated among different people in organizations is collected and synthesized in the forecasting process. Therefore, prediction markets are an innovative tool for discovering, visualizing and generalizing knowledge about a specific topic among employees in organizations through crowdsourcing. The price in the prediction market is a bid for a specific event. It also represents the approximate cost that the bidder assigns to the parameters considered in the bid. 1.1 Terminology A market maker is the participant on the market that plays the role of a liquidity provider; artificial intelligence, which, through a mathematical algorithm, contributes to the smooth flow of concluded transactions and thereby provides instant liquidity. The market valuation rule is a function S (F, y) that depends on the distribution function F and the values of the predicted value y. It is envisaged that under this rule, points are calculated, reflecting the quality or success of the forecast [4]. If F1 , . . . , FN are the series of realizations of predictive distribution functions, and y1 , . . . , yN are the series of actually observed values of a forecasted value, then the prediction is more effective [5], the higher the average point calculated as. 1 N S(Fi , yi ) i=1 N

(1)

Prediction market is a platform that gathers and aggregates knowledge and judgment of a large, diverse group around a particular event or concept to generate forecasts. Predications represent the group’s overall knowledge and quantification of the likelihood of future results. The purpose of the prediction market is to forecast the likely success of a product, idea, or political candidate [6].

272

O. Olefir and O. Lastovsky

2 Description of the Algorithm Prediction markets represent a variety of markets that typically use a continuous double auction (CDA) structure, but they suffer from low liquidity and, as a consequence, do not reveal their predictive power. Automated trading setting up liquidity required by the market is one of the solutions to the problem of insufficient number of users. Bringing automated market makers on board can be very beneficial for the prediction markets [7]. The most popular automated market marker, used in the digitized prediction markets, is the logarithmic market scoring rule (LMSR) [4]. However, this rule, which is the most practical to use, has a number of significant drawbacks, including possible losses of funds by the market maker and the manual regulation of market liquidity [8]. The task is to modify the existing logarithmic market valuation rule so that our market maker automatically adjusts the price depending on the agility that is being observed, and also to minimize the loss of funds of the market maker in the worst-case scenario. LMSR is based on costs function:  q   i , (2) exp C(q) = b log i b where q—the vector of the net amount of sold shares. b > 0—the liquidity parameter accountable for the marginal loss of the market maker. The cost function is used to determine the value of a transaction as follows: If you take q1 as the quantity of sold shares before the transaction and q2 after that, then the transaction value v is calculated as follows: v = C(q2 )−− C(q1 ). The worst-case scenario for a market maker will happen if every participant, except the market maker himself, knows exactly which event of n options will happen. At the beginning of the auction, considering that q1 = q2 = . . . = 0, then C(0, 0, . . .) = b ∗ ln(n), and in the worst-case   scenario, considering that q1 is a correct = q1 . Therefore, the market receives response, C(q1 , −∞, . . .) = b ∗ ln exp qb1 (q1 − b ∗ ln(n)) monetary units and its lost represents (q1 − b ∗ ln(n) − q1 ) = b∗ln(n). Because the cost function is nonlinear, the price of shares is not scalable according to the number of shares sold. However, the cost function is differentiated, so the limit price can be set for number k of shares:   exp qbi C(q)  , = (3) p(q) = qi exp qk k

b

that in the context of a prediction market can be interpreted as an estimate of the likelihood that an event of the relevant share will occur. Modification of the rule is in additional dependence of the net amount of sold shares on the cost function and a corresponding modification of the pricing formula. The cost function will be as follows: 

  qi exp C(q) = g(b, q) ∗ log , (4) g(b, q) i

Application of a Modified Logarithmic Rule for Forecasting

where g(b, q) = b−1 ∗



273

qi , b > 0 – dependence on the number of shares in

i

circulation for accelerating the growth of price in the dominance of a potentially correct forecast, which would allow reduce losses in the worst-case scenario for the market maker. In this case, the pricing rule becomes more complex [9]: ⎛ ⎞

 q j ⎠ pi (b, q) = b−1 ∗ log ⎝ exp g(b, q) j     (5)   qj qi − q exp q exp j j j j g(b, q) g(b, q)   +   qj j qj j exp g(b, q) To illustrate the impact of this modification on the work of the market maker, a program was written in python that reproduced the logic of the prediction market. For simplicity, only binary variants of markets with an initial uniform distribution were considered. The coefficient b of 20 applies for all cases (the modified rule is adapted by the input parameters to the classical one with integration and implementation in already existing system).

Fig. 1. The classical rule, b = 20

Figures 1–2 illustrate the relationship between q1 and p1 at a fixed q2 = 250, 500 and 750 respectively [9]. The increasing price slope with increasing q2 on the modified rule chart indicates the liquidity sensitivity of the market maker.

274

O. Olefir and O. Lastovsky

Fig. 2. Modified rule, b = 20

Figures 3–4 show the value of shares in case of alternation of purchase of each share one by one. As can be seen, the ratio b = 20 corresponds to a low level of liquidity, but under increasing price, in the case of a modified rule, it is scaled up according to the number of participants. Since the price is initially high enough, with the purchase of the maximum number of shares under a potentially correct forecast, the market maker losses will be much less than in the case of the classic rule, and the subsequent transactions will guarantee a profit to cover the costs (in such case the price is optimized for a larger audience and the accuracy of the forecast will be improving).

Fig. 3. Modified rule

It should be noted that the same input data are used under modified rule, which makes the transition from standard LMSR easier, without significant changes to the program code.

Application of a Modified Logarithmic Rule for Forecasting

275

Fig. 4. The classic rule

3 Estimation of Predictive Efficiency Regression methods were chosen to compare the performance of PM forecasts using the Bitcoin cryptocurrency price prediction method. In addition to conventional linear regression [10], two main ensemble methods for two main types of meta-algorithms were used, namely: bagging and boosting, represented by Random Forest and Gradient Boosting. Ensemble methods are a paradigm of machine learning where several models (often called "weak learners") are taught to solve the same problem and are merged to obtain better results. The main hypothesis is that, under the right combination of weak models, more accurate and reliable models can be produced [11]. Random Forest is an ensemble machine learning regression method that combines bootstrap and bagging techniques to reduce the retraining problem and increase the precision compared to a single tree. Prediction is as a result of aggregating the responses of many trees. Training for each tree is independent of each other (on different subsets), supporting solving the problem of building identical trees on the same data set and making this algorithm convenient for use in the system of distributed computing [12].; Gradient Boosting is an ensemble method (representative of a booster method) that reduces the task to gradient descent. It is a machine learning technology for regression and classification that creates a prediction model in the form of an ensemble of weak prediction models, usually the decision trees. The model is built in stages, as other boosting methods do, and generalizes them, allowing to optimize a differentiated loss function [13]. To compare the effectiveness of the prediction market performance, it is necessary to obtain user trade data as to the specific event that can be predicted by different mathematical methods [14–20]. However, such data cannot be retrieved from the network, since it is not publicly available, and the existing services of this nature do not provide relevant data. In addition, there are no prediction markets that simultaneously implement CDA and market maker as a method of data aggregation, and therefore, comparing these

276

O. Olefir and O. Lastovsky

methods for one event at the same time interval can only be done based on own data, which needs to be collected. Therefore, it was necessary to create separate prediction market system with cloud hosting and cloud computing and attracting real people to trade with real funds. A software system was built to collect prediction market data that was user-friendly and intuitive for ordinary users. The main code that implements the logic of the prediction market is written in python.

4 Experiment In order to determine the effectiveness of the prediction market performance, Bitcoin cryptocurrency price fluctuations were collected and recorded by the created system every minute during the period of two weeks, as well as trade data related to exchange between the users themselves and the market maker within the prediction market. Three criteria were used to assess the quality of the constructed models: • MS — mean square error; • R2 — coefficient of determination; • SR — success rate, the percentage of correct answers. Table 1 shows the resulting forecast at the time of the event using Linear Regression (LG), Random Forest (RF) and Gradient Boosting (GB) 10 min before the event, relative to the price of the cryptocurrency Bitcoin at this moment (Price). Table 2 shows the average indicators of the goodness of fit of the models we built for forecasting the next 10 min. Table 3 shows the initial data of the prediction market, that is, the price of the shares “Yes”, for which the auction ended 10 or 60 min before the event. MSR – this is the probability, in percentage, of the occurrence of the event using the market maker mechanism of aggregation of the prediction market user data, and CSR - respectively, using the CDA. The results obtained using different prediction market mechanisms with the same questions are quite similar. However, there is a greater categoricalness of the users in the event of an event in the case of a market maker with a modified logarithmic market valuation rule, in the case of both time values. Figure 5 shows the accuracy of binary prediction of Bitcoin cryptocurrency using regression methods and different types of prediction market, where MLMSR is a modified logarithmic market valuation rule, LG is a linear regression, RF is a random forest method, GB is a gradient boosting method. The findings indicate that a market maker with a modified logarithmic market valuation rule and a CDA generally perform better than forecasts using regression methods. In some cases, the CDA performs worse than the Random Forest and Gradient Boosting methods. This is likely due to a liquidity problem as a result of a lack of the market maker on the prediction market while using the CDA mechanism. Since for a rapid response to the exchange rate fluctuation, news or insider information, the immediate ability to sell or buy shares is required, the market maker provides

Application of a Modified Logarithmic Rule for Forecasting

277

Table 1. Price forecast 10 min before the event Date

LG

RF

GB

Price

2019-07-29

9508.89

9497.99

9507.26

9519.77

2019-07-30

9568.30

9574.82

9564.66

9533.48

2019-07-31 10030.73 10028.92 10020.16 10030.52 2019-08-01 10451.56 10452.17 10457.69 10411.13 2019-08-02 10418.20 10401.38 10405.75 10404.48 2019-08-03 10873.39 10872.53 10877.40 10885.88 2019-08-04 10970.35 10975.05 10978.58 10944.11 2019-08-05 11721.95 11771.47 11737.05 11727.29 2019-08-06 11629.30 11645.40 11645.21 11648.05 2019-08-07

9508.89

9497.98

9507.26 11868.32

2019-08-08 11588.95 11599.33 11589.75 11593.43 2019-08-09 11835.62 11928.08 11851.40 11865.04 2019-08-10 11304.09 11345.70 11337.01 11316.58 2019-08-11 11394.47 11362.72 11385.62 11403.53

Table 2. Goodness of fit of constructed models Tool R2 (10) MSE(10) LG

0.86

30.2

RF

0.73

41.3

GB

0.76

39.0

this opportunity (while in its absence, users have to wait for processing of the purchase or sale application). That is, until there is another user who agrees to make the transaction under the stated terms, the user does not have the ability to influence the price and, accordingly, the likelihood of an event. At the same time, the market maker is ready to buy the shares from the user at the market price, and in case of too high probability and accordingly the price, in the opinion of the user, there is always an opportunity to buy the shares of the opposite option to the question by the market for the price at an acceptable value (since the user is confident about the event more compared to current probabilities of the prediction market).

278

O. Olefir and O. Lastovsky Table 3. The resulting probabilities of prediction market Date

CSR(10) CSR(60) MSR(10) MSR(60)

0

2019-07-29 67

81

82

75

1

2019-07-30 92

73

81

98

2

2019-07-31 90

84

91

95

3

2019-08-01 66

96

81

83

4

2019-08-02 36

39

49

61

5

2019-08-03 83

87

80

92

6

2019-08-04 89

77

66

75

7

2019-08-05 88

72

77

92

8

2019-08-06 39

41

19

45

9

2019-08-07 66

72

64

58

10 2019-08-08 49

37

30

23

11 2019-08-09 52

64

88

65

12 2019-08-10 21

15

7

9

13 2019-08-11 80

78

86

74

Fig. 5. Accuracy of binary prediction 10 and 60 min prior to the event

5 Conclusions The paper analyzes the mathematical methods of forecasting in a market environment, investigates the principles of the prediction market operation and finds that the use of PM as a predictive tool is appropriate; and, depending on the possibilities of liquidity and number of active users, different methods of data aggregation should be used.

Application of a Modified Logarithmic Rule for Forecasting

279

A market maker using a modified logarithmic market valuation rule is the optimal solution in a low liquidity environment only if the PM is used only as a forecasting tool, and the result is of paramount importance as the system will suffer losses when using the market maker. In a case when the liquidity is not an issue for the PM, the CDA user aggregation mechanism should not be worse than the PM with a market maker, since users will be able to do both – to sell shares immediately and also to have more accurate price values since the market maker using the modified market valuation rule changes stock prices with some leaps to mitigate the worst-case scenario, and therefore this mechanism is more optimal for PM, envisaged for long-term perspective. A key advantage of the prediction market over regression methods is that the determination coefficient in the case of regression methods decreased significantly with increasing interval to the time of the event occurrence, together with the accuracy of the forecast, while the PM provided fairly accurate forecasts even within a large interval prior to the event occurrence. The CDA prediction market, however, is still a viable option in the conditions of insufficient activity by the users and is a better method than forecasting with application of regression methods within the sufficiently long prediction intervals. The obtained data differ from the already existing data since they are obtained in real conditions of trade by users with two fundamentally different mechanisms of aggregation of data of trading in shares at the same time, which allows to compare their effectiveness with one another and with mathematical methods for which data were collected within the same time span. The value of the results obtained is that the proposed methods and tools make it possible to identify the best possible way of forecasting developments in the field of finance. The developed mathematical and software for obtaining and multilevel data processing of the forecasting market significantly simplify the analysis of the data, contribute to the validity of making various decisions. The results of the study can be used to develop more advanced models of the prediction market, to make forecasts in finance or other industries.

References 1. Armstrong, J.S.: Forecasting for Marketing. Quantitative Methods in Marketing, pp. 92–119. International Thompson Business Press, London (1999) 2. Tikhonov, E.E.: Forecasting in market conditions. Nevinnomyssk, 221 p. (2006) 3. Rhode, P., Strumpf, K.: Historical presidential betting markets. J. Econ. Perspect. 18(2), 127–142 (2004) 4. Hanson, R.: Logarithmic market scoring rules for modular combinatorial information aggregation. J. Predict. Mark. 1(1), 1–15 (2007) 5. Armantier, O., Treich, N.: Eliciting beliefs: proper scoring rules, incentives, stakes and hedging. Eur. Econ. Rev. 62, 17–40 (2013) 6. Spann, M., Skiera, B.: Internet-based virtual stock markets for business forecasting. Manag. Sci. 49, 1310–1326 (2003) 7. Schervish, M.J.: A general method for comparing probability assessors. Ann. Stat. 17(4), 1856–1879 (1989)

280

O. Olefir and O. Lastovsky

8. Murphy, A.H.: A new vector partition of the probability score. J. Appl. Meteorol. 12(4), 595–600 (1973) 9. Othman, A., Pennock, D.M., Reeves, D.M., Sandholm, T.: A practical liquidity-sensitive automated market maker. ACM Trans. Econ. Comp. 1(3), 25 (2013). Article 14 10. Locasto, M.E., Wang, K., Keromytis, A.D., Salvatore, J.S.: FLIPS: hybrid adaptive intrusion prevention. In: Valdes, A., Zamboni, D. (eds.) Recent Advances in Intrusion Detection. LNCS, vol. 3858, pp. 82–101 (2005). https://doi.org/10.1007/11663812_5 11. Hoerl, A.E., Kennard, R.W., Hoerl, R.W.: Practical use of ridge regression: a challenge met. J. R. Stat. Soc. Ser. C 34(2), 114–120 (1985) 12. Hastie, T., Tibshirani, R., Friedman, J.: Chapter 15. Random Forests. The Elements of Statistical Learning, pp. 587–623 (2009) 13. Mason, L., Baxter, J., Bartlett, P.L., Frean, M.: Boosting algorithms as gradient descent. In: Solla, S.A., Leen, T.K., Müller, K. (ed.) Advances in Neural Information Processing Systems 12, pp. 512–518. MIT Press (1999) 14. Lytvynenko, V., Kryvoruchko, O., Lurie, I., Savina, N., Naumov, O., Voronenko, M.: Comparative studies of self-organizing algorithms for forecasting economic parameters. Int. J. Modern Educ. Comput. Sci. (IJMECS) 12(6), 1–15 (2020). https://doi.org/10.5815/ijmecs. 2020.06.0 15. Bodyanskiy, Y.V., Tyshchenko, O.K., Deineko, A.O.: An evolving neuro-fuzzy system with online learning/self-learning. IJMECS 7(2), 1–7 (2014). https://doi.org/10.5815/ijmecs.2015. 02.01 16. Sajedi, H.: Elham Masoumi,"Construction of high-accuracy ensemble of classifiers. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 6(5), 1 (2014). https://doi.org/10.5815/ijitcs.2014.05.01 17. Ganguly, K.K., Siddik, Md.S., Islam, R., Sakib, K.: An environment aware learning-based selfadaptation technique with reusable components. Int. J. Modern Educ. Comput. Sci. (IJMECS) 11(6), 53–64 (2019). https://doi.org/10.5815/ijmecs.2019.06.06 18. Gupta, D.K., Goyal, S.: Credit risk prediction using artificial neural network algorithm. Int. J. Modern Educ. Comput. Sci. (IJMECS) 10(5), 9–16 (2018). https://doi.org/10.5815/ijmecs. 2018.05.02 19. Khazaee, S., Faez, K.: A novel classification method using hybridization of fuzzy clustering and neural networks for intrusion detection. IJMECS 6(11), 11–24 (2014). https://doi.org/10. 5815/ijmecs.2014.11.02 20. Abernethy, J., Chen, Y., Vaughan, J.W.: An optimization-based framework for automated marketmaking. In: Proceedings of the ACM Conference on Electronic Commerce (EC) (2011)

A Survey on Kernelized Fuzzy Clustering Tools for Data Processing Zhengbing Hu1

and Oleksii K. Tyshchenko2(B)

1 National Aviation University, Liubomyra Huzara Ave 1, Kyiv 03058, Ukraine 2 Institute for Research and Applications of Fuzzy Modeling, CE IT4Innovations,

University of Ostrava, 30. dubna 22, 701 03 Ostrava, Czech Republic

Abstract. Although most of the well-known fuzzy clustering algorithms are somewhat sensitive to noise, there are some more profound possibilistic fuzzy clustering techniques based on kernel distance metrics that are not subject to this problem. The paper presents a survey on this type of fuzzy clustering methods. Meanwhile, introducing a particular distance measure based on the Bregman divergence to a fuzzy clustering tool made it possible to improve the algorithm’s performance compared to traditional Euclidean-based analogs. A bunch of experimental evaluation is performed for this set of methods. Keywords: Kernel-based fuzzy clustering · Distance measure · Feature vector · Sample weighting · Kernel function

1 Introduction Clustering endeavors to identify a concealed structure in data patterns. Clustering is a requisite and valuable instrument to examine the data. It is extensively practiced to solve queries in numerous fields, such as biology [1], computer vision [2], marketing [3], recommender systems [4], etc. The necessary task of clustering is maximizing the data similarities within the same cluster while concurrently minimizing data correspondence coming from distinct clusters. As the idea of partial memberships is founded by Zadeh [5] and employed to clustering by Ruspini [6], the fuzzy clustering scheme draws substantial attention due to its performance and simplicity [7–10]. The most regularly employed fuzzy clustering procedure is the highly reputable fuzzy c-means (FCM) to have been offered originally by Dunn [11] and further broadened by Bezdek [12, 13]. Fuzzy clustering further explicates notable competitive capacity in the data exploratory analysis because it can flexibly assign data items to clusters. Studying the hidden formation of real data and the crucial role of membership vectors in classification, numerous fuzzy clustering methods have been recommended to improve memberships for a better understanding of diverse data [14–16]. In the last few years, some additional fuzzy clustering schemes [17] were suggested by concentrating on the study of decent distance computation between data units and cluster centers. In fuzzy clustering, membership vectors are applied to illustrate the attribution of data objects to a particular cluster (each data point obtains a membership vector to express its belongingness © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 281–289, 2021. https://doi.org/10.1007/978-3-030-80472-5_24

282

Z. Hu and O. K. Tyshchenko

to various clusters). When admitting membership as a novel exhibition of the original data, its subspace structure should be geometrically compatible with the initial data. Accordingly, it is supposed that some regularization routines dependent on the reconstruction terms, which are widely used in self-representation methods, can be adapted to restrain the membership rates, thereby enhancing the representation of fuzzy clustering. Instinctively, all the traditional structure-preserving regularization tools [18, 19] can be employed to control the subspace formation of the raw data in the novel description of memberships. Nevertheless, due to the stochastic restriction of memberships (the sum of memberships for each data unit equals one), it is customarily challenging to optimize the goal function of fuzzy clustering in the context of the before-mentioned regularization penalties. A considerable bunch of examinations has revealed that kernel functions can make the scheme more robust to noise and outliers [20, 21]. As an unsupervised dynamic optimization process stated in [13], FCM requests to iterate the membership level and cluster center repeatedly until the procedure converges, hence a high load of samples diminishes the real-time appearance of this procedure. When a bunch of individuals relating to various classes in the data collection are quite diverse, FCM divides part of units in the class with more samples into the class with fewer individuals, which produces severe misclassification of representations and the degradation of clustering execution. The intention of this inquiry is to make a survey of the currently applicable kernel-based FCM modifications based on the Bregman divergence instead of the classical Euclidean distance and to provide a thorough comparison of their implementation efficiency with the traditional versions of the fuzzy clustering tools. The manuscript’s structure indicates the key stages of the study issues. A concise introduction on the FCM primary information is manifested in Sect. 2. Section 3 sheds some light on preliminaries of the Bregman divergence. A brief summary on kernel functions is described in Sect. 4. A survey of enhanced kernel-based fuzzy clustering procedures is mentioned in Sect. 5. Eventually, some experimental investigations are granted in Sect. 6. Lastly, Sect. 7 covers the conclusions.

2 FCM Baseline Information FCM [8–10] is a nonlinear iterative optimization scheme, which advances a combination of the K-means clustering and the fuzzy set theory. The chief principle of the method is to keep to a minimum value the cost function by renewing a class centroid and membership values in a repeating manner. Items are ascribed to various groups depending on the membership levels in order to reach a clustering goal. The FCM’s objective function is expressed as detailed below JFCM =

N M i=1

k=1

m xi − ck 2 wki

subject to 0 ≤ wki ≤ 1, i = 1, 2, . . . , N ; k = 1, 2, . . . , M ;

A Survey on Kernelized Fuzzy Clustering Tools

M k=1

0
0, πi (ai ) = 1 ai ∈Ai

a∈A

In contrast to Nash equilibrium, the Best Response method (BR) forms the optimal strategy of the agent in response to the actions of all other agents. The corresponding operator of the value of the state of the system in method (6) has the form [10]: ⎞ ⎛ L       Qti (a) πj aj ⎠ BR Qti = max⎝ πi

a∈A

j=1

Pareto Equilibrium (PE) takes place in a Common-Interest Markov Game, where the winning matrices are the same for all players Qti (a) = Qti (a)∀i, j ∈ I , ∀a ∈ A [11]. A game with different payoff matrices can turned into a game with common L L     k  λk Qt (a) πjPE aj , where λj > interests through a convolution PE(Qi ) = k=1

j=1

a∈A

0(j = 1 . . . L). The search for a PE solution of the game is carried out by an independent choice of agent strategies, similar to the search for a NE solution. Multi-agent play is optimal for Pareto, if there is no strategy of players,  common  which allows to improve the winnings of all players: Qti π PE ≥ Qti (π ).   Pareto-optimal mixed strategies π PE = π1PE , . . . , πLPE can obtained by maximizL L     k  λk Qt (a) πj aj → max. ing the convolution of concave (up) gain functions: k=1

a∈A

j=1

π

+ 1) in (6) based on optimal collective strategies π ∗

=   ∗ π1 , . . . , πL∗ (NE, PE, etc.) requires considerable computational work. For practical applications, it is sometimes sufficient to maximize payment functions. Then, instead of (6), we use a modified method of estimating the matrices of gains [12]:   i Qt+1 (8) (a) = (1 − αt )Qti (a) + αt rti + γ max Qti (a) . Calculating the value of V i (t

a

Matrix Stochastic Game with Q-learning for Multi-agent Systems

309

For ensure the convergence of method (8), it is necessary to impose restrictions on the rate of change of its adjustable parameters. The general restrictions are as follows: ∞ ∞   αt = ∞, αt2 < ∞, where αt > 0 is positive sequences of real quantities

t=0

t=0

decrease monotonically. The parameter learning step can calculated as follows: αt = α0 t −κ ,

(9)

where α0 > 0 is the initial value of the parameter αt ; κ ∈ (0, 5; 1] is the order of learning speed of the method (8). The choice of solutions under uncertainty is based on a random distribution with probabilities proportional to the values of the functions Qi (a). The probability of choosing the i-th agent of the action ai (k) can be determined based on the Boltzmann distribution: ∗

πi (ai (k)) =

eQi (ai (k))/ T , k = 1 . . . Ni , Ni  ∗ (a (j)) T Q i / i e

(10)

j=1

where T is the temperature parameter of the system; Qi∗ (ai (k)) = max r i (a−i , ai (k)), a−i ∈ A−i , A−i =

L

×

j=1,j=i

Aj ;

Ni 

a−i

πi (ai (k)) = 1. For large T values, a close to uniform

k=1

random distribution is realized, and for small T values is a distribution close to the “greedy” choice of agents’ actions is realized, when the action with the largest Q-value is most often chosen. Choosing of pure strategies ⎧

⎫ ⎛ ⎞

k ⎨ ⎬   

π i ai (j) > ω⎠, k = 1 . . . Ni ∀i ∈ I , ai = Ai (k)

k = arg⎝min (11) ⎩ ⎭ k

j=1 i  the ibasis of random distributions based on mixed strategies π =  i is carried iout on π [1], . . . .π [Ni ] ∈  , ∀i ∈ I where ω ∈ [0, 1] – is a real random number with a uniform distribution. The convergence of the method is estimated by the error of the condition of additional non-rigidity [13], weighted by mixed strategies:  πi − π˜ i 2 ,

= L−1 (12) i∈I

 where πi = diag(πi )∇Vi V i ; diag(πi ); is formed from vector elements Ni , formed    from vector elements πi ; ∇V i = V i j |j = 1 . . . Ni is vector function of average Ni    V i j πi j is the function winnings for fixed net strategies of the i-th player; V i = j=1

of the average winnings of the i-th player; ∗ is Euclidean norm of the vector. The complementary non-rigidity condition describes the solutions of the Nash game in equalizing mixed strategies. The considered condition additionally takes into account the solutions of the game in pure strategies. A qualitative indicator of the convergence

310

P. Kravets et al.

of the game is the growth of the functions of discounted total winnings ϒi (2) or the function of winnings, averaged over the number of players: ϒ = L−1

L 

ϒi ,

(13)

i=1

3.3 Stochastic Game Solving Algorithm 1. Set the parameters of the game: Lis number of players;  Ni is the number of strategies of each player i = 1 . . . L; Ai = ai (1), . . . , ai (Ni ) is sets of discrete actions (pure   strategies) of players; vi (a) ∀a∈A is matrix of winnings of a deterministic game; initial time t = 0; initial values of stochastic game payoff matrices: Qti (a1 , . . . , aL ) = ε, ∀ai ∈ Ai , ∀i = 1 . . . L, where 0 < ε 0 is temperature parameter of the system. 2. Calculate the values of mixed strategies π = (π1 , . . . , πL ) based on current estimates of payoff matrices Qti (a1 , . . . , aL ) ∀i = 1 . . . L according to (10). 3. Make a random selection of agents a = (a1 , . . . , aL ) based on strategies in to (11). accordance π = (π1 , . . . , πL ) according  4. Obtain current agent winnings rt = rt1 , . . . , rtL according to (1). 5. Calculate the parameter of the learning  istep: αt accordingto (9). 6. Modify the payoff matrices Qt+1 = Qt+1 (at )|i = 1 . . . L according to (8). 7. Calculate the error of fulfillment of the condition of additional non-rigidity (12) and value of  the  the function of average gains ϒ (13). i 8. If Qt+1 − Qti  < ε ∀i = 1 . . . L, then specify t := t + 1 and go to step 2.   9. Output the calculated values of payoff matrices Q = Q1 , . . . , QL and strategies π = (π1 , . . . , πL ). End.

4 Experiments Let’s solve the stochastic game of two agents (L = 2) with two pure strategies (N = 2). The matrices of average winnings take the values: (0.4; 0.2); (0.5; 0.9)) - for the first player; ((0.9; 0.1); (0.1; 0.9)) - for the second player. Wins dispersions take the same value d (a) = d = 0, 01 ∀a ∈ A for all players. For solve the game problem, perform evaluation of Q-functions by method (8) with the parameters: T = 0, 1; α0 = 1; κ = 0.7; γ = 0.5. The initial approximations of the elements of the payoff matrices take the same values Q(a) = 0, 01 ∀a ∈ A. Mixed game strategies are determined by the Boltzmann distribution (10). Sections of the functions of the average gains V i shown in Fig. 1. As can see in Fig. 1, the game has two Nash solutions in pure strategies: (π1 [1], π2 [1]) = (0; 0), (π1 [1], π2 [1]) = (1; 1)[1]) = (1; 1) and one solution in mixed strategies: (π1 [1], π2 [1]) = (0, 5; 0, 67).

Matrix Stochastic Game with Q-learning for Multi-agent Systems

311

Fig. 1. Functions of average gains of agents a) first agent, b) second agent and c) deviation of the game trajectory from the optimal solution for T = 0, 5

If one player’s net strategy is fixed at Nash’s equilibrium point, no change in the other player’s strategy can improve his winnings. The discrepancy of the method for T = 0, 5 is illustrated by the deviation of the learning trajectory of the game from the optimal value (0, 0) in Fig. 1c.

5 Experimental Results of Computer Simulation Graphs of the functions of average gains ϒ and norms of deviation of mixed strategies from their target values are given in Fig. 2a on a logarithmic scale. The decrease in the graph of the error rate of the condition of additional non-rigidity (12) indicates the convergence of the game Q-method to the Nash equilibrium point. The value of the temperature coefficient T has a significant effect on the convergence of the game Q-method. The rate of convergence is determined by the steepness of the decay of the graph of the function . The tangent of the acute angle formed between the linear approximation of this graph and the time axis can estimate the order of the convergence rate. As the value of T increases, the order of the rate of convergence of the game Qmethod decreases. As can see in Fig. 2a, close to 1 order of convergence rate of method (8) is provided for the values T ∈ (0; 0, 2].

Fig. 2. a) Characteristics of the convergence of the game Q-method and b) dependence of the average number of steps of learning a game on the parameter γ .

312

P. Kravets et al.

Let’s study the dependence of the time of convergence of the game Q-method on its parameters. The required training time is defined as the minimum number of steps required to reach the Nash equilibrium point: tout = (t = tmin | t ≤ ε ), where ε is given training accuracy. The average number of steps of training agents is determined during a kexp  1  series of experiments with different sequences of random variables: ¯t = kexp tout j , j=1

where kexp is number of experiments. The results are averaged for kexp = 1000 experiments of learning the stochastic game with accuracy ε = 10−3 . In Fig. 2b is shown a graph of the average number of steps of learning a stochastic game from the value of the discount parameter γ of current winnings (2), obtained for the following parameters of the method: κ = 0, 7; T = 0, 1. The increase of γ leads to a slowdown in the reduction of discounted current winnings and, consequently, to a decrease in the average number of steps in learning a stochastic game.

6 Discussion The rate of convergence of the game method is determined by the order κ of the rate of decrease of the parameter αt , that determines the current value of the learning step of the method (8). The dependence of the average number of steps of learning a stochastic game on the parameter κ is given in Fig. 3a. Data obtained for the Q-method with the parameters: γ = 0, 5; T = 0, 1. With increasing of κ number of steps required to learn the game Q-method with accuracy ε = 10−3 , increases. The influence of the variance of current winnings on the convergence of the stochastic game with Q- learning is shown in the form of a graph in Fig. 3b.

Fig. 3. Dependence of the average number of steps of learning a game on the a) parameter κ and b) winnings dispersion

The result of growing the dispersion d of current winnings is an increase in the number of steps required to Q-learning a stochastic game.

7 Conclusions Performed research confirms the iterative method of Q-learning with restrictions on its parameters provides a solution to the stochastic game in conditions of payoff matrices uncertainty, and it can be used to support the collective decision-making.

Matrix Stochastic Game with Q-learning for Multi-agent Systems

313

The practical use of this method is limited to compliance with the conditions of convergence to one of the states for the collective equilibrium. In the conditions of game uncertainty, the values of the parameters that ensure the fulfillment of convergence conditions can be established theoretically based on the results of the theory for stochastic approximation, or experimentally in the course of computer modeling. In this paper, the ranges of changes in the parameters of the game Q-method are experimentally established to ensure the convergence to one of the Nash equilibrium points. As the value of the current win discounting parameter increases, the variance of current winnings reduces, and the order of change for the learning step decreases, and the convergence rate of the game Q-method increases. The results of a computer experiment confirm the convergence of the game algorithm under the constraints of stochastic optimization. The probability of the obtained results is confirmed by the repeatability of the results of a computer experiment for different sequences of random variables. The disadvantage of stochastic play is the low (with a degree order) rate of convergence, due to the process of training players in conditions of a priori uncertainty. The positive point is the possibility of parallelizing the stochastic game with the use of powerful computing tools to accelerate the process of its convergence for practical applications. New research results in this direction can be expected from the introduction of other criteria for the formation of player payments and other methods of teaching stochastic play, for example, using artificial intelligence methods.

References 1. Neyman, A., Sorin, S.: Stochastic Games and Applications. (vol. 570). Springer Science & Business Media, Berlin (2003). https://www.springer.com/gp/book/9781402014925 2. Fudenberg, D., Drew, F., Levine, D.K., Levine, D.K.: The Theory of Learning in Games. vol. 2. MIT press, Cambridge (1998). ISBN 9780262061940 3. Weiss, G.: Multiagent Systems: A Modern Approach to Distributed Artificial Intelligence. MIT press, Cambridge (1999). ISBN 9780262232036 4. Wooldridge, M.: An Introduction to Multiagent Systems. John Wiley & Sons, Hoboken (2009). ISBN 978-0-470-51946-2 5. Hashemi, A.B., Meybodi, M.R.: A note on the learning automata based algorithms for adaptive parameter selection in PSO. Appl. Soft Comput. 11(1), 689–705 (2011). https://doi.org/10. 1016/j.asoc.2009.12.030 6. Watkins, C.J., Dayan, P.: Q-learning. Mach. Learn. 8(3–4), 279–292 (1992). https://doi.org/ 10.1007/BF00992698 7. Kaelbling, L.P., Littman, M.L., Moore, A.W.: Reinforcement learning: a survey. J. Artif. Intell. Res. 4, 237–285 (1996). https://doi.org/10.1613/jair.301 8. Sutton, R.S., Barto, A.G.: Reinforcement Learning: An Introduction. Stanford University Press, Redwood City (2018). ISBN 978-0-262-19398-6 9. Hu, J., Wellman, M.P.: Nash Q-learning for general-sum stochastic games. J. Mach. Learn. Res. 4(Nov), 1039–1069 (2003) 10. Weinberg, M., Rosenschein, J.S.: Best-response multiagent learning in non-stationary environments. Proc. Third Int. Joint Conf. Auton. Agents Multiagent Syst. 2, 506–513 (2004) 11. Podinovskii, V.V., Nogin, V.D.: Pareto-Optimal Solutions of Multicriteria Problems. Nauka, Moscow (1982).(in Russian)

314

P. Kravets et al.

12. Chen, H.F.: Stochastic Approximation and its Applications. (vol. 64). Springer Science & Business Media, Berlin (2006). https://www.springer.com/gp/book/9781402008061 13. Moulin, H.: Game Theory with Examples from Mathematical Economics: Transl. from French. Moskow, Mir. (1985). (in Russian) 14. Burov, Y., Vysotska, V., Kravets, P.: Ontological approach to plot analysis and modeling. In: CEUR Workshop Proceedings, pp 22–31 (2019). Electronic copy: http://ceur-ws.org/Vol2362/paper3.pdf 15. Kravets, P., Lytvyn, V., Vysotska, V., Ryshkovets, Y., Vyshemyrska, S., Smailova, S.: Dynamic coordination of strategies for multi-agent systems. In: Babichev, S., Lytvynenko, V., Wójcik, W., Vyshemyrskaya, S. (eds.) ISDMCI 2020. AISC, vol. 1246, pp. 653–670. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-54215-3_42 16. Loganathan, M., et al.: Reinforcement learning based anti-collision algorithm for RFID systems. Int. J. Comput. 18, 155–168 (2019) 17. Singh, S., Trivedi, A., Garg, N.: Collaborative anti-jamming in cognitive radio networks using Minimax-Q learning. Int. J. Mod. Educ. Comput. Sci. (IJMECS) 5(9), 11–18 (2013). https:// doi.org/10.5815/ijmecs.2013.09.02 18. Salukvadze, M.E., Beltadze, G.N.: Stochastic game with lexicographic payoffs. Int. J. Mod. Educ. Comput. Sci. (IJMECS) 10(4), 10–17 (2018). https://doi.org/10.5815/ijmecs.2018. 04.02 19. Dembri, A., Redjimi, M.: Towards a meta-modeling and verification approach of multi-agent systems based on the agent petri net formalism. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 11(6), 50–62 (2019). https://doi.org/10.5815/ijitcs.2019.06.06

Topology Synthesis Method Based on Excess De Bruijn and Dragonfly Heorhii Loutskii, Artem Volokyta(B) , Pavlo Rehida, Artem Kaplunov, Bohdan Ivanishchev, Oleksandr Honcharenko, and Dmytro Korenko National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv 03056, Ukraine [email protected]

Abstract. Scaling high performance computer systems needs increasing the fault tolerance at the design stage of a topology. There are several approaches of designing simple fast routing with fault tolerance. One of effective approach is to ensure fault tolerance at the topology level. This article discusses methods for synthesizing fault-tolerant topologies. These topologies are hierarchical, scalable with a simple routing system. This paper discusses the synthesis of a multilevel topological organization consisting of clusters of excess De Bruijn rank connected by Dragonfly-like connections with hierarchical scaling. The main characteristics of this topology are high fault tolerance at the cluster level, as well as ultra-high scaling speed with a linear increasing of topology degree. The issue of routing in this kind of system is considered, the comparison with other topological organizations is made. Keywords: Dragonfly · De Bruijn · Multilevel topological organization

1 Introduction and Related Work One of the key factors that determines the parameters of a distributed system is its topology. The system of connections between the elements determines the speed of message transmission, and the complexity of routing, and fault tolerance, and the price of the system. Different topologies provide different advantages, but have their disadvantages. The problem of finding the optimal topology does not have an unambiguous solution, and the reason is the mutual exclusivity of the requirements. The topology must provide high fault tolerance; have acceptable topological characteristics; provide fairly simple and efficient routing; scale well and maintain the acceptability of scalable characteristics, but at the same time provide a wide range of system size choices. As a result, the problem of topology synthesis, both for a specific problem and in general, is relevant [1–6]. Special attention should be paid to several issues when developing a topology. First of all, it is a question of fault tolerance. Modern high-performance systems contain a large number of nodes (up to 109 ), as a result, the failure rate is quite high. In the context of topologies, the issue of fault tolerance is primarily a question of fault tolerance routing. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 315–325, 2021. https://doi.org/10.1007/978-3-030-80472-5_27

316

H. Loutskii et al.

Another important issue is the issue of routing. Ease of routing directly affects the speed of message transmission. Universal methods, such as tabular routing, are optimal for systems with a dynamic topology, but they are not a good choice in the case of a static system. It is good when you can search for neighbors of a certain node, knowing only its number. This reduces the cost of routing memory. Even better, when the search for each subsequent forwarding address can be performed through simple conversions over the number, because in this case the router can be implemented in hardware. The main objective of this research is to propose method of synthesis fault tolerant topologies with next properties: – Simple and fast fault tolerant routing; – Ultra-high scaling speed with a linear increasing of topology degree. The key methodologies, used on this research, are a graphs theory and practis of fault tolerant routing. The structure of paper is presented. Introduction highlightes actuality of this research, the objective and methodology are shown. In 2nd section the synthesis of multilevel topology based on Excess de Bruijn clusters is described. In 3rd section considered the proposed method of topologies synthesis. 4th section is devoted to the main parameters of topologies, created by this method, are shown. The 5th section it is conclusion. The main features of these are highlighted.

2 Synthesis of Multilevel Topology Based on Excess De Bruijn Clusters 2.1 Setting Objectives for Synthesis of Multilevel Topology Excess De Bruijn The main things to consider at this stage are the principles of networking and scaling. As mentioned above, one of the key tasks in synthesis is to ensure good scalability. This topology is an extension of the usual de Bruijn topology [7–10] using a excess binary number system (containing an additional digit -1, denoted as T) [11–15]. All vertices of the topology are encoded in this system, and connections are formed by shifts to the left and right with the insertion of different digits (thus, each node has 6 neighbors). This topology of rank 2 is shown in Fig. 1. Another important task is the simplicity of routing. A tiered system is assumed to contain at least two components: a base cluster of the excess de Bruijn topology and external Dragonfly-like connections [2]. These two topologies have a number of significant differences, which in turn are bad for routing. Thus, it is necessary to develop the logic of synthesis so that it can be used in the construction of the route. The proposed method allows to meet both of these requirements, providing an easy way of synthesis and rapid growth of system performance during scaling. Another feature is that it works identically for any node coding systems and any clusters. This makes it possible to consider the idea and the resulting topology separately from each other. A simple construction scheme allows you to use offsets in routing: all the neighbors of each node can be found through offset operations on the code, and the route search performed through the step-by-step insertion of the destination code in the source code.

Topology Synthesis Method Based on Excess De Bruijn and Dragonfly

317

11

10

T1

01

00

0T

1T

T0

TT

Fig. 1. Excess de Bruijn rank 2

Another routing option has been tree-based routing, which allows you to bypass failures. Also a feature of this topology is the presence of vertices with different code, but the same number, which allows you to perform fault substitution and redistribution of roles. 2.2 Method of Synthesis of Multilevel Topology This section considers the idea of synthesis as such. As an example, a line of two processors was chosen: the simplest version of the cluster. This will allow you to take several steps of scaling, avoiding the explosive growth of the number of vertices and, as a consequence, display problems. This cluster is shown in Fig. 2. To increase performance, the idea of hierarchical scaling is proposed: • For a system of rank R, a cluster is a system of rank R-1. • The system contains as many clusters as it contains vertices in the previous step. Consider a system of rank 2. In the system of the previous rank 2 vertices, and therefore, it will contain 2 clusters, numbered as 0 and 1. In Fig. 2 shows a given system. Thus, each node will have 2 numbers: internal (inside the cluster) and external (number of the cluster itself). Denote them as a (external number) and b (internal number). As a result, the full address of the node A = a.b. The formation of external connections is performed according to the following algorithm: to get a neighbor, the numbers a and b in the full address of the node are swapped. That is: A = a.b

(1)

adest = b, bdest = a

(2)

Adest = adest bdest = b.a

(3)

318

H. Loutskii et al.

Example: node 01. Each number (internal and external) is described by one binary digit. That is, a = 0, b = 1. Then, according to the described algorithm: adest = b = 1, bdest = a = 0

(4)

Adest = adest .bdest = b.a = 1.0

(5)

Thus, knowing the principle of forming an external number and the principle of routing within a cluster, you can easily find the path between any two vertices of the topology. The following algorithm is offered: 1. Let Asrc = a.b, Adest = c.d 2. Perform routing within the source cluster. a.b → a.c 3. Transfer between clusters. a.c → c.a 4. Perform routing within the destination cluster. c.a → c.d If b = c, a = c or a = d, the corresponding steps are skipped. To improve this algorithm, you can apply a few more points: • If a = c (i.e., both vertices in the same cluster) then intercluster routing is not involved at all. • If the direct intercluster connection is broken, a bypass is applied. The first address to bypass is cluster d. That is, the full route will look like this:   a.b → a.c → c.a fail → a.d → d .a → d .c → c.d If this path does not work - use the search for paths through different clusters. Full route (x - address of the bypass cluster):   a.b → a.c → c.a fail → a.x → x.a → x.c → c.x → c.d Thus, the fault tolerance of intercluster routing will be equal to the number of clusters (the algorithm will try to list all clusters and perform a bypass through each). The length of any intercluster bypass will be constant, equal to L = 3 ∗ Dcluster + 2. We move on to the next step of scaling. The system of the previous rank becomes a new cluster, and the number of these clusters is equal to the number of vertices in the cluster. That is, the system of rank 3 contains 4 clusters of 4 vertices. This is shown in Fig. 2. Regarding numbering, the system is as follows: the internal number of the node becomes its full number from the previous step, and the external number is the number

Topology Synthesis Method Based on Excess De Bruijn and Dragonfly

0

1 0

0

0 Cluster 00

1 0

1

0 Cluster 01

0

1

1 0

1

0

1

0

1 1

Cluster 10

1 0

0 Cluster 11

319

1 1

System

Fig. 2. System for example, rank 3

of the subsystem, which includes this node. In other respects, the principle of connection remains the same, only the bit size of the parts of the number changes. As for the routing algorithm, it also remains unchanged. The only nuance is that the intracluster routing algorithm will now be the routing algorithm in the previous rank system. As a result, this routing can be considered as recursive. Example. Let the source be the node 1000, the receiver - 1101. 1. Routing within the cluster 10: 1000 → 1011. a. Routing within a 2: 00 → 11 rank topology i. Transfer within the cluster 0: 00 → 01 ii. Transfer between clusters: 01 → 10 iii. Transfer within the cluster 1:10 → 11 2. Transfer between clusters: 1011 → 1110 3. Transfer within cluster 11: 1110 → 1101 a. Routing within a topology of rank 2:10 → 01 i. Transfer within cluster 1: not required ii. Transfer between clusters: 10 → 01 iii. Forwarding within cluster 0: not required

320

H. Loutskii et al.

After performing 2 scaling steps, you can display the system characteristics. Each of them can be described by a recurrent formula, and from it - to derive the usual formula for rank. It is known that the number of vertices at each step increases by a number of times equal to the number of vertices in the previous step. From here: 4 = N12 NR+1 = NR2 = NR−1

R

(6)

You can also see that a maximum of 1 external connection is added to each node of the system at each step. Therefore: SR+1 = SR + 1 = S1 + R

(7)

Now about the diameter. At each step, routing is divided into 3 steps: routing within the sending cluster, direct forwarding between clusters, and routing within the destination cluster. Intracluster routing is described by the same algorithm, but for a smaller order system. So: =

2R D1

DR+1 = 2DR + 1 = 4DR−1 + 2 + 1 =  i R R R + R−1 i=0 2 = 2 D1 + 2 − 1 = 2 (D1 + 1) − 1

(8)

Here we can highlight the following feature: the faster the number of nodes N relative to the growth of the topological characteristics of SD the better. This means that the system increases its performance faster than the incidental costs and data transmission delays. Thus, the primary task for scalability is to provide such a switching system and such scaling that as the rank of the topology increases, the SD/N ratio goes to zero.

3 DragonDeBruijn System Applying this scaling principle to the excess cluster de Bruijn, we can obtain the topology shown in Fig. 3. One of the interesting aspects of the DragonBruijn network is excess clusters. These are clusters that duplicate each other due to the same number, but different code. In this case, such clusters are pairs 01-1T and 0T-T1. If we consider the nodes of the system only in the context of routing, it means that there is a so-called quasi-quantum transition between these nodes. Therefore, even if an entire cluster fails, the compute vertices hidden inside it will still be available. Similarly, this makes it possible, in addition to the usual routing, to use the so-called “quasi-quantum routing” [15]. Its essence is to send between vertices with the same number but different code. Due to this, the bypass can be performed even in cases where conventional routing methods do not provide such an opportunity. Another advantage is that such duplications exist not only within clusters and between entire clusters, but also between single vertices scattered across the topology. This makes the quasi-quantum transition convenient not only for bypassing failure, but also for transitioning between distant vertices or clusters. This would in some cases reduce the actual transferring time. Knowing the features of excess code, you can try to calculate the number of such “quantum portals”. It is known that the excess code contains patterns.

Topology Synthesis Method Based on Excess De Bruijn and Dragonfly

321

11

10

T1

01

00

0T

1T

T0

TT

11

11

10

T1

11

10

01

00

0T

1T

T1

0T

T0

TT

10

T1

01

00

0T

1T

T0

TT

10

11

01

00

01

11

10

1T

T1

00

0T

T0

TT

11

10

01

1T

T1

0T

T0

TT

T1 10

T1

0T

1T

11

10

01

00

1T

T0

TT

00

11

01

00

1T

T1

01

00

0T

T0

TT

1T

T0

TT

0T

T0

11

10

T1

01

00

0T

1T

T0

TT

TT Fig. 3. DragonDeBruijn system, rank 2

If there is a pattern P-Q in the coding system, then for any code of the form aPb the equality is valid: aPb = aQc

(9)

The bit size of parts a and c does not matter. Currently, the following 2 patterns are selected for the 2-digit excess code: 01-1T and 0T-T1 [15]. Knowing them, you can determine the type and number of vertices that are duplicated. Table 1 lists such vertices. By looking at the connections between the clusters, you can see that they are uneven. The main communication centers are clusters 00, 10 and T0. Clusters 01, 11, T1, 0T, 1T, TT are “connected” to them. There is no connection between the “chains” through the hidden vertices. But there is a passage between the chains: 01 → 1T (between the “chains” 0 and 1) and 0T → T 1 (between the “chains” 0 and T). This means that with the complete failure of the entire external communication system (ie, in the case where the clusters are completely unconnected) due to quasi-quantum tunnels, you can restore connectivity by performing transitions on the line TT - T0 - T1 - 0T - 00 01 - 1T - 10 - 11. Another good news is that the clusters in the “chain” are connected by 3-fold quasi-quantum bonds, and the interchain transitions are connected by 9-fold (each node of one cluster is connected to the corresponding node of the other). As a result, it significantly reduces the length of intra-cluster transitions and provides good fault tolerance.

322

H. Loutskii et al. Table 1. Hidden quasi-quantum links

№

Node 1

Node 2

Transcluster link

№

Node 1

Node 2

Transcluster link

2

0010

01T0

00 ↔ 01

-2

00T0

0T10

00 ↔ 0T

3

0011

01T 1

-1

00T1

0T11

1

001T

01T T

-3

00TT

0T1T

10

1010

11T 0

6

10T0

1T10

11

1011

11T 1

7

10T1

1T11

9

101T

11T T

5

10TT

1T1T

-6

T010

T1T 0

-10

T0T0

TT10

-5

T011

T1T 1

-9

T0T1

TT11

-7

T01T

T1TT

-11

T0TT

TT1T

10 ↔ 11

T0 ↔ T1

10 ↔ 1T

T0 ↔ TT

Substituting the initial parameters of the excess de Bruijn in the formula, you can get the following parameters: R

4 = N12 = 92 NR+1 = NR2 = NR−1

R

(10)

SR+1 = SR + 1 = 5 + R

(11)

DR+1 = 2DR + 1 = 4DR−1 + 2 = 2 ∗ 2R + R = 2R+1 + R

(12)

  SDR+1 = SR+1 ∗ DR+1 = (5 + R) 2R (D1 + 1) − 1

(13)

This method of scaling allows achieving an explosive increase in the number of vertices with a linear increase in degree. In this case, the excess coding in the node numbers is stored at all stages of scaling. As a result, all the benefits of redundancy de Bruijn associated with fault tolerance remain.

4 Results In addition to the standard degree, diameter and multiplicative SD index, several other indicators are proposed for estimating topologies. The first is the ratio of the SD index to the number of vertices. This indicator will allow to estimate deterioration of characteristics of topology concerning its size. For example, if two topologies have the same SD but different number of vertices, the one with more nodes is better. Conversely, with the same number of vertices, a topology with a lower SD is preferred. For convenience, denote it as J. J =

SD N

(14)

Topology Synthesis Method Based on Excess De Bruijn and Dragonfly

323

Table 2. Topological characteristics Rank

1

2

3

4

5

6

Topology

Param.

Hypercube

N

2

4

8

16

32

64

S

1

2

3

4

5

6

D

1

2

3

4

5

6

SD

1

4

9

16

25

36

J

0.5

1

1.125

1

0.78

0.56

M

–

1.5

1.25

0.875

0.56

0.34

N

3

9

27

81

243

729

S

2

5

6

6

6

6

D

1

2

3

4

5

6

SD

2

10

18

24

30

36

J

0.66

1.11

0.66

0.29

0.12

0.04

Excess De Bruijn

Dragonfly-like

Dragon DeBruijn

M

–

1.33

0.44

0.11

0.03

0.01

N

2

4

16

256

65536

429496729

S

1

2

3

4

5

6

D

1

3

7

15

31

63

SD

1

6

21

60

155

378

J

0.5

1.5

1.3125

0.23

0.002

8.801E–08

M

–

2.5

1.25

0.1625

0.0014

5.19E–08

N

9

81

6561

43046721

1.85E+15

3.43E+30

S

5

6

7

8

9

10

D

2

5

11

23

47

95

SD

10

30

77

184

423

950

J

1.11

0.37

0.011

4.27E–06

2.28E–13

2.76E–28

M

–

0.27

0.007

2.48E–06

1.28E–13

1.53E–28

Table 2 shows the characteristics of topologies for different ranks. As it shown in Table 2, realizations, includes from excess versions has higher degree of topology. It allows to increase fault-tolerance. The second proposed indicator is the ratio of the SD-index increment to the increment of the number of vertices. It will allow to estimate deterioration of characteristics of topology with scaling. That is, the better the topology scales in terms of characteristics (the weaker the degree and diameter increase compared to the increase in productivity), the smaller this indicator will be. M =

SD N

(15)

324

H. Loutskii et al.

As shown in Table 2, the proposed structures, although average for SD, but differ in low J and M with increasing rank. Topologies based on hypercube and Dragonfly-like have a standard degree. Unlike this, Excess De Bruijn have stable degree, and DragonDeBruijn has a little more degree, but has additional benefits for high scalable. However SD parameter increases too fast, that’s why it is required to research Cartesian product of Excess De Bruijn and Dragon DeBruijn rank 2 or 3. The important aspect is scalability. On the one hand, other things being equal, the topology that provides the best growth in the number of nodes is better. On the other hand, excessive growth rates require the development of additional methods to limit scaling. At the same time, special attention should be paid to topological characteristics, as they tend to deteriorate in the process of node building. There are many ways to synthesize scalable topologies with the desired characteristics, but each has its own characteristics and limitations.

5 Conclusions In this paper, it is proposed to consider the modified method of synthesis based on combination of topologies. The topology of excess de Bruijn rank 2 was used as a basic cluster. This topology was proposed and investigated in previous publications, its advantages are high fault tolerance, simple routing and good topological characteristics. It is recommended to use a Dragonfly-like communication system to connect clusters. Each node of each cluster connects to no more than one node of another cluster in such a way that each cluster has at least 1 connection to each other cluster. The developed topology in comparison with others has very interesting topological characteristics. It shows fast growth number of nodes with increasing rank. Also in this topology, unlike others, there is a rather small degree and excellent routing. Quite interesting is the method of combining several topologies into one multilevel structure. This allows you to combine the advantages of both structures, but often creates problems: the deterioration of the resulting topological characteristics, the difference in the principle of node coding, the complexity of routing, and others. Avoiding these problems is a separate task that also needs to be addressed. Thus, using this topology as a basis immediately allows you to solve a number of issues related to fault tolerance. Similarly, the availability of ready-made simple routing methods allows you to close the issue of routing at the cluster level.

References 1. Alverson, R., Roweth, D., Kaplan, L.: The gemini system interconnect. In: 2010 18th IEEE Symposium on High Performance Interconnects, pp. 83–87. IEEE (2010) 2. Kim, J., Dally, W.J., Scott, S., Abts, D.: Technology-driven, highly-scalable dragonfly topology. In: 2008 International Symposium on Computer Architecture, pp. 77–88. IEEE (2008) 3. Ajima, Y., Sumimoto, S., Shimizu, T.: Tofu: A 6D mesh/torus interconnect for exascale computers. Computer 11, 36–40 (2009)

Topology Synthesis Method Based on Excess De Bruijn and Dragonfly

325

4. Guan, K.C., Chan, V.W.S.: Cost-efficient fiber connection topology design for metropolitan area WDM networks. J. Opt. Commun. Netw. 1(1), 158 (2009). https://doi.org/10.1364/JOCN. 1.000158 5. Hu, Z., Tyshchenko, O.K.: The possibilistic gustafson-kessel fuzzy clustering procedure for online data processing. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) Advances in Computer Science for Engineering and Education III. ICCSEEA 2020. Advances in Intelligent Systems and Computing, vol. 1247, pp. 419–428. Springer, Cham (2020). https://doi.org/10. 1007/978-3-030-55506-1_38 6. Mukhin, V., et al.: Method of restoring parameters of information objects in a unified information space based on computer networks. Int. J. Comput. Netw. Inf. Secur. 12(2), 11–21 (2020). https://doi.org/10.5815/ijcnis.2020.02.02 7. Ganesan, E., Pradhan, D.K.: The hyper-debruijn networks: Scalable versatile architecture. IEEE Trans. Parallel Distrib. Syst. 4(9), 962–978 (1993) 8. Dürr, F.: A Flat and Scalable Data Center Network Topology Based on De Bruijn Graphs (2016). arXiv preprint arXiv:1610.03245. 9. Kamal, M.S., Parvin, S., Ashour, A.S., Shi, F., Dey, N.: De-Bruijn graph with MapReduce framework towards metagenomic data classification. Int. J. Inf. Technol. 9(1), 59–75 (2017). https://doi.org/10.1007/s41870-017-0005-z 10. Peng, G., Ji, P., Zhao, F.: A novel codon-based de Bruijn graph algorithm for gene construction from unassembled transcriptomes. Genome Biol. 17, 232 (2016). https://doi.org/10.1186/s13 059-016-1094-x 11. Olexandr, G., Rehida, P., Volokyta, A., Loutskii, H., Thinh, V.D.: Routing method based on the excess code for fault tolerant clusters with infiniband. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) Advances in Computer Science for Engineering and Education II. ICCSEEA 2019. Advances in Intelligent Systems and Computing, vol. 938. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-16621-2_31 12. Honcharenko, O., Volokyta, A., Loutskii, H.: Fault-tolerant topologies synthesis based on excess code usign the latin square. In: The International Conference on Security, Fault Tolerance, Intelligence ICSFTI2019, Ukraine, Kyiv, 14–15 May 2019, pp. 72–81 (2019) 13. Loutskii, H., Volokyta, A., Rehida, P., Goncharenko, O.: Using excess code to design faulttolerant topologies. Tech. Sci. Technol. 1(15), 134–144 (2019). https://doi.org/10.25140/ 2411-5363-2019-1(15)-134-144 14. Loutskii, H., Volokyta, A., Rehida, P., Honcharenko, O., Thinh, V.D.: Method for synthesis scalable fault-tolerant multi-level topological organizations based on excess code. In: Hu, Z., Petoukhov, S., Dychka, I., He, M. (eds.) ICCSEEA 2020. AISC, vol. 1247, pp. 350–362. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-55506-1_32 15. Loutskii, H., Volokyta, A., Rehida, P., Honcharenko, O., Ivanishchev, B., Kaplunov, A.: Increasing the fault tolerance of distributed systems for the Hyper de Bruijn topology with excess code. In: 2019 IEEE International Conference on Advanced Trends in Information Theory (ATIT), Kyiv, Ukraine, 2019, pp. 1–6 (2019). https://doi.org/10.1109/ATIT49449. 2019.9030487

Method for Cyber Threats Detection and Identification in Modern Cloud Services Zhengbing Hu1,2 , Sergiy Gnatyuk2,3,4(B) , Berik Akhmetov3 , Volodymyr Simakhin4 , Dinara Ospanova5 , and Nurbol Akatayev6 1 Central China Normal University, Wuhan, China 2 National Aviation University, Kyiv, Ukraine

[email protected] 3 Yessenov University, Aktau, Kazakhstan 4 International Research and Training Center for Information, Technologies and Systems,

Kyiv, Ukraine 5 Kazakh Humanitarian Juridical Innovative University, Semey, Kazakhstan 6 Satbayev University, Almaty, Kazakhstan

Abstract. Today cloud technologies and their applications are implementing in various ICT infrastructures. It has led to increased attention to the problems of cyber threats, the growth of which is inseparably linked with the growth of ICT. In this paper the analysis of the existing models, systems and methods for cyber threats detection was carried out for their disadvantages defining. A model of cloud service has been developed; it allows to ensure the security of cloud service based on cloud computing and conduct appropriate simulations. Improved method for cyber threats detection has been developed, it allows to detect cyber threats in cloud services and classify them. The developed method was experimentally investigated using NSL-KDD data base as well as simulation tools RStudio and CloudSim. It was proved the correctness of its work and the possibility of application in cloud services as well as increase efficiency of cloud system security by 48.02%. In addition, a cloud service model has been developed that can be used to build cloud services based on the various cloud computing architecture. In the future, based on the proposed method and model, appropriate tools for detecting and classifying cyber threats in cloud services can be developed. It is significant because it can be autonomous functional unit of SIEM as well as other instrumental tools of CSIRT/SOC. Keywords: Information technology · Cloud Service · Information security · Cyber threat · Detection · Identification · NSL-KDD

1 Introduction The usage of cloud computing has gained a significant advantage due to the reduced cost of ownership of IT applications, extremely fast entry into the services market, as well as rapid increases in employee productivity [1]. Everything can be implemented in the cloud service: from data storage to data analysis, applications of any scale or size. Employees © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 326–346, 2021. https://doi.org/10.1007/978-3-030-80472-5_28

Method for Cyber Threats Detection and Identification

327

also implement their own cloud applications for work, contributing to the development of their own cloud culture (BYOC). In addition, the use of cloud services is now available not only for large enterprises, but also for companies in medium and small businesses, which makes cloud technologies one of the main environments for the operation of their information systems [2]. However, such an increase in the efficiency of working with cloud technologies has led to increased attention to the problems of cyber threats, the growth of which is inseparably linked with the growth of ICT [3]. A cloud service user can deploy their own applications, build their infrastructure, or simply process data, but in any case, they trust their confidential data to the cloud service provider and want to be sure that their data is secure. Providing information security (IS) in a cloud environment is the responsibility of the provider, and therefore their systems must meet a number of requirements of both national and international law and international recommendations. Therefore, the main scientific and technical problem can be formulated as follows: data security may be compromised and there is a risk of mass data loss by many users due to the possibility of conducting cyber threats in cloud services. Because information is not only stored in the cloud, but is also processed, users must be confident in the security and availability of their data. The solution to this problem can be provided by using various methods of cyber threat detection (MCD), IDS/IPS systems, cyber incident response modules, etc. Cyber threat is any circumstance or event that may cause a breach of information security policy and/or damage to an automated system [4–6]. The main purpose of cybersecurity is to prevent the implementation of existing cyber threats, which are the sources of the following risks [7–10]: 1) Loss of intellectual property; 2) Violation of compliance and regulations; 3) Compromising credentials and authentication; 4) API threats; 5) Hacking accounts; 6) Improper usage of cloud services. In cloud services cybersecurity ensuring is very actual and important challenge today. From this position, this research study is significant and proposed MCD will be useful for cybersecurity monitoring in real cloud systems. This paper consists of the following sections: 1) Introduction; 2) Review of up-to-date methods for cyber threats detection and problem statement; 3) Theoretical background of method development; 4) Experimental study and discussion; 5) Conclusions.

2 Review of Up-to-Date Methods for Cyber Threats Detection and Problem Statement 2.1 Literature Review Cloud computing systems have a multi-level architecture of different services and levels of management. Figure 1 depicts a classification of data security threats on each layer of the cloud system. Security issues for the SaaS platform can generally be divided into two categories: attacks on development tools and attacks on management tools. In general, all threats can be divided into three groups: 1) threats to data confidentiality; 2) attacks on the interface; 3) SSH attacks. Security issues for the IaaS and PaaS platforms are grouped into four classes: attacks on cloud services, attacks on virtualization, attacks on unified computing, and attacks on SLAs [11, 12].

328

Z. Hu et al.

In the Table 1 shows a multi-level classification of threats for the three layers of clouds, which are the first level.

Fig. 1. Classification of data security attacks on each layer of the cloud system

At the next level are cloud services, and at the third - the types of attacks on these services [13]. There are many MCDs, but they all use very similar techniques for direct detection. A significant disadvantage of most of them is that they are described only theoretically and have no practical confirmation (verification). 1)

2)

Method for cyber threat recognition based on fuzzy feature clustering [14, 15]. The essence of splitting objects that contain signs of anomalies (cyberattacks) into a class of sets of the same type in terms of information security or cybersecurity lies in splitting sets of objects into subsets. The method involves the usage of machine learning. Traditional fuzzy clustering algorithms use a given number of partition clusters as input parameters, and some of them also use a given cluster fuzzy index in the space of signs of vulnerabilities, anomalies, NSD threats and cyberattacks [16, 17]. Based on the information criterion of functional efficiency for the IP, a mechanism for adjusting the parameters of the algorithm for clustering threat signs can be implemented. Method for detecting cyber threats using Big Data technology [18]. Data management and expertise methods such as biometric authentication protect against cyberattacks by providing security solutions to the massive protection of data volumes. By analyzing Big Data logs, we can prevent cyber threats by monitoring data. When Big Data analysis is combined with JIT (Just in Time) analysis, it collects information on machines that have an open connection to locations outside the LAN. It also predicts future attacks and provides information about previous attacks that may have taken place on your system. IBM has developed a big data solution that protects data from threats and fraud. IBM’s solution detects risk and intrusion when analyzing structured and unstructured data. QRadar correlates in real time, detects anomalies and reports for immediate threat detection, and sends

Method for Cyber Threats Detection and Identification

329

Table 1. Multilevel classification of data security threats in cloud computing Layer of cloud (service)

Cloud service

Security threat

Attack type

Risk value

SaaS

Web service

Data security

Confidentiality

Medium

Interface attack

Signatures attack

Low

Attacks on users credentials

Medium

API

IaaS and PaaS

Virtualization platform

SSH attacks

Hardware level virtualization

Software level virtualization Development services

3)

4)

5)

Attacks on API keys Medium Attacks on users credentials

Medium

ARP spoofing on virtual switching

High

MAC spoofing on virtual switching

High

Hacking on computing

Low

Cloud software

Harmful software Scripts

High

Computing services

Unified Attacks during data computing attacks processing

Low

SLA attacks

High

Hacking

rich security data to IBM Big Data products such as IBM InfoSphere BigInsights. Large datasets need to be reduced to successfully detect anomalies. Method for detecting cyber threats using the analysis of social networks [19]. Among the main classes of methods used in ACC, we can distinguish the following: methods of graph analysis, statistical methods, data mining, methods of optimization theory and algorithm theory. It is also convenient to single out the methods of semantic analysis and text analysis. In this case, it is necessary to verify whether the system supports the language in which users of the analyzed social network communicate. These methods are used to identify the following major threats: network spam; threats to social engineering; password theft and phishing; web attacks; leakage of information and compromising the behavior of company employees; Advanced Persistent Threat (APT) attacks. Method for cognitive security using artificial intelligence [20]. It is the concept of a self-defense network that identifies a potential threat on the Internet and takes appropriate action to prevent “confidential data from being compromised”. At the same time, you use a combination of a number of modern technologies to identify and analyze key threats (both external and internal to the client) using special techniques for analyzing real-time data behavior. Method for detecting cyber threats using a structured behavioral model [21]. The method is based on the analogy of comparing natural language and network traffic. First, the Trace Sequence of the captured network traffic is determined with the

330

Z. Hu et al.

parameters: active text, active grammar, active vocabulary, and ratio. The next step is to start processing the captured sequence using DBSCAN – a spatial data clustering algorithm where noise is present. Next, the data from the analysis of the captured piece of traffic is compared with the normal behavior of traffic that was obtained in an isolated environment, and is called the reference. Having certain differences results in detection of cyber threats. 6) Method for “deep analytics” [22, 23]. This method is a combination of popular and effective methods (predictive analytics, descriptive analytics, graph analysis, analysis of unstructured information, optimization), which together give the desired result - the detection of cyber threats or anomalies. Examples of method implementation: detection of anomalies; statistical threat profiles; relationship analysis. 7) Method for detecting cyber threats of Yu. Malachi [24]. The system and method of detecting a cyber threat in accordance with implementations of the present invention include automatic detection of resources in the network, resource detection unit, emulation, fake asset creation unit, at least one resource detected in the network, associating a trap sensor with malware with emulation resource and detection by malicious trap sensor, malware related to the emulated resource. The system and method may also include downloading data related to the detected malware on the server, analyzing the downloaded data on the server to obtain the analysis result and performing one or more actions based on the analysis result. 8) Radar Services method for detecting and counteracting cyber threats [25]. The method is based on the use of: many systems for analysis of signature and behavioral analysis of network traffic and next-generation isolated software technologies for analysis of all incoming e-mails; advanced correlation engine that analyzes network traffic and events that use statistical models, recursive methods and machine learning to differentiate normal and abnormal behavior and identify patterns; the usage of a risk and safety management team that analyzes, verifies and aggregates all findings. 9) Network streaming and threat detection approach [26]. Numerous studies have been conducted on real-time threat detection systems. The problem of processing large amounts of data on network traffic of corporate systems, while providing real-time monitoring and detection, was considered, which remain unresolved. In particular, they introduced and evaluated a flow-based threat detection system that can quickly analyze overly intensive real-time network traffic data using streaming flow-based clustering algorithms to detect abnormal network actions. 10) SANS company cyber threat detection, prevention and control system [27]. This system is described only theoretically (in the form of recommendations), and contains 5 aspects: use of security measures based on end-to-end visibility; avoid excess information; use security solutions that perform real-time analysis; reduce latency within the network; introduction of “deep” protection. 11) Method for data collection, processing, analysis and storage for monitoring cyber threats and their notification to users [28]. The system collects intelligence data from multiple sources and then pre-processes the intelligence data for further analysis by the intelligence analyst. The analyst reviews the intelligence and determines if it is appropriate for the client to sign a cyber threat alert service. The system

Method for Cyber Threats Detection and Identification

331

reforms and collects intelligence data and automatically transmits intelligence data through many delivery methods. 12) Method for planning the structure of cyber threats and their application to reduce the impact of cyber threats [29]. A security system consisting of a computer, memory, data storage, containing a dictionary of the intellectual capacity of cyber-threats and a technological dictionary; and an application stored in memory. When executed by a computer, the program generates a report that identifies the intent of the cyber threat and identifies the cyber threat technology, in which the intent of the cyber threat is selected from several intentions for the cyber threat listed in the dictionary of cyber threat imposition and in which cyber threat technology is selected from the technology dictionary. 2.2 Problem Statement After the analyzing of known MCDs [14–29], it is clear that not all of these methods have been qualitatively and experimentally studied. Almost half of them have high requirements for computing resources and are not easy to implement, and therefore are described only theoretically. In addition, due to the technology of the method itself, not all algorithms have the ability to log new, not yet assigned to any category of cyber threats. Also, only fuzzy feature-based clustering, Cognitive Security Method, Network Streaming and Threat Detection System, MCD of SANS company, and MCD using cyber threat planning have real-time cyber threat detection. In the review of the literature for each method was not said that the study was also conducted in cloud computing systems, but indicated the possibility of such implementation for the methods: MCD based on fuzzy clustering of features, MCD using Big Data technology, the method of “deep analytics” and Network streaming and threat detection system. In general, this analysis indicates the problem of detecting cyber threats in cloud environments of any type and services of any type. The main purpose of this work is to develop a method for detecting cyber threats in cloud services. To achieve this purpose we need to solve the following tasks: 1. Develop a model of cloud service and based on it a MCD to ensure the security of cloud services by further neutralizing the identified threats; 2. Experimentally investigate the MCD to verify its correct operation and the possibility of application in cloud services.

3 Theoretical Background of Method Development 3.1 Technological Architecture of Secure Cloud Service Based on Cloud Computing Technology Cloud environment in which the MCD will be introduced in this section. The technology architecture is based on the recommendations of Cisco, which has developed its own progression of evolution of cloud data centers: 1) consolidation and aggregation of data center assets; 2) abstraction, is a key phase, because the assets of the data center are

332

Z. Hu et al.

abstracted from the services that are actually supplied; 3) automation, which is capitalized on consolidated and virtual aspects, fast backup services and automatic modeling; 4) the interaction of the corporate “cloud” with the public; 5) the final phase – “inter-cloud”, which replaces the existing types of clouds. Before building the architecture of the cloud data center, it is necessary to identify the components of the data center blocks in the basis of cloud architectures.

Fig. 2. Technological architecture of the data center based on cloud computing

10 Gigabit Ethernet. The data center is designed with a high density of virtual machines that are combined with a large number of processors. From a network perspective, the growth of virtual machines and the concentration of cores will facilitate the transition to 10 Gigabit Ethernet as a necessary mechanism for providing servers. Specific benefits

Method for Cyber Threats Detection and Identification

333

of the transition include: real-time policy configuration; mobile security and network policy, which is replaced by the policy of the virtual machine during its mobility; continuous operation of management models that establish management and operation of the environment for virtual machines and physical servers. Unified Fabric, which gives all servers (physical or virtual) access to the local network, storage network, and IPC network, allowing them to be more integrated into the customer’s network to increase efficiency and cost savings. Unified Computing. The unified structure allows you to fully virtualize a “cloud” data center with pools of computer, network, and storage resources using unified computing. Unified Computing covers silos in a classic data center, allowing more efficient use of infrastructure in a fully virtualized environment, and creates a single architecture using standard technologies that ensure compatibility and investment protection. The Unified Computing system combines computing and network capacity, storage system access and virtualization resources in a scalable modular design that is managed as a single energy-saving system. This system can be managed using the built-in control system, in the Unified Computing platform. Figure 2 shows the technological architecture, which presents the “cloud” data center of the next generation. The diagram shows examples of component blocks for the data center. In general, the completed architecture contains not only components of the structure, but also is regulated by different types of service and regulatory requirements. The architectural model offers 9 tiers of the data center network: application software; virtual machine and distributed virtual switch (virtual machine, VSwitch); storage and storage networks (storage, SAN); calculation (compute); access; aggregation; core, where there is also a module for detecting cyber threats; peering; basics of the Internet (IP-NGN backbone).

3.2 Information Security System Architecture for the Protected Cloud Service Model Based on Cloud Computing Technology Along with the technological component of the architecture of data centers, an important place is also occupied by the issue of trust in the infrastructure model of “cloud” computing [25]. The key to gaining an advantage from the cloud is to establish a trust approach that begins with the establishment of such attributes in cloud architecture. Trust in a “cloud” data center is based on several basic concepts: 1) Security: traditional data issues and resource access control, encryption and incident detection; 2) Control: the ability of the enterprise to directly manage the processes of deployment of applications; 3) Compliance and maintenance at the management level: compliance with general requirements; 4) Timely detection of cyber threats, prevention of intrusions, blocking cyberattacks. Figure 3 shows the structure of a protected data center based on Cloud Computing technology from the point of view of security, namely the models of threats and measures that need to be taken to minimize risks. The structure also reflects full control, compliance with requirements and agreements on the level of services.

334

Z. Hu et al.

The key idea of this model is that information security should not be secondary or simply part of the overall security, it should be disseminated and implemented at all levels of the architecture. The threat profile consists of such elements as: service disruption; data leakage; data disclosure; data modification; identity theft and fraud; intrusion. As can be seen from Fig. 3 detection of cyber threats is one of the most important tasks in the system of information security of cloud environments.

Fig. 3. The structure of a secure data center based on cloud computing

Control. Aspects of security control of the “cloud” data center vary from management of systems support to management of access control systems. As a first step, you should review the basic level of security to more rigid conditions. Usually, the centralization of data leads to more frequent internal threats. Therefore, the compartmentalization strategy is the most key component of data control. Moreover, unencrypted data in the data center should be considered as part of risk management policy. Compliance and Service Level Agreement. Aspects of Compliance and SLA possess quite multifaceted features. It is necessary to clearly understand the features of the various components of the “cloud”, which must be compatible. All service level agreements should be synchronized with key assurance mechanisms. 3.3 Groups of Cyber Threats and Cyberattacks in Cloud Environments One of the most complete descriptions of all cyber threats and attacks that can be implemented is in the KDD database [30]. NSL-KDD is a data set proposed to solve some of

Method for Cyber Threats Detection and Identification

335

the integral problems of the KDD’99 data set mentioned. Although this new version of the KDD dataset is still not without some problems, and is not the best guide to existing real networks, and due to the lack of available datasets for network identification systems, it is used as an effective reference dataset, which will help researchers compare different MCD. In addition, the number of entries in NSL-KDD sets and test sets is processed. This advantage makes it available to run experiments without having to accidentally pick a small portion. Thus, the results of the evaluation of different research papers will be consistent and comparable. The NSL-KDD dataset has the following advantages over the initial KDD dataset: 1) It does not include redundant entries in the data set, so classifiers will not be biased for a recurring entry. 2) There are no duplicate records in the proposed test sets; therefore, productivity is not biased by methods that have better detection rates. 3) The number of selected records from each complex group is inversely proportional to the percentage of records in the original KDD dataset. As a result, the classification indicators of different teaching methods change in a wider range, which makes it more effective to accurately assess different teaching methods. 4) The number of entries in the set and test sets is clear, making it possible to experiment on a complete set without having to randomly select a small part. 3.4 Block Diagram of the Proposed Method for Detecting Cyber Threats Figure 4 shows a block diagram of detecting cyber threats in a cloud environment. When the host is connected to the cloud environment, network traffic begins to be generated. Next is the data processing unit (data process), where the network traffic arrives at the behavior analyzer (behavior analyzer), which contains the records of the NSL-KDD database. The analyzer compares the data captured from the network traffic with the database and begins to use classifiers to determine. The next block to which the data is transmitted is the block of identification and analysis, where the pre-classified threat is analyzed in detail on certain grounds, and it is determined to which elements the threat was directed. After that, the system issues a warning message that part of the traffic is abnormal, and at the same time begins to check the identified threat with previously found or recorded immediately in the database. If such a match is found, a notification with further actions can be issued (in case of their previous successful application). The last two modules are actually a record of detailed data about the threat (date, group, whether it was previously identified, etc.) and the formation of a mini-report for review, which shows the overall result. Taking into account the characteristics of cloud computing systems and ideal cyber threat detection systems, the developed method meets the following requirements: 1. Processing of large-scale dynamic multilevel autonomous computing systems and data processing environments. Clouds are large-scale systems based on virtual machines that are automatically created, transferred, and removed at the user’s request at runtime. It is generally assumed that the middleware provider initially

336

Z. Hu et al.

reported changes in resources, but in cloud computing involving large networks and systems, it is important to automatically support these changes without human intervention. To overcome the complexity of its dynamic nature, the process of detecting cyber threats must be able to cope with it without human intervention, which facilitates the monitoring and control of network elements in real time. 2. Identify various attacks with the least false positives. Due to the growing number of attacks, their complexity and unpredictability, the system must recognize new attacks and their vulnerable intentions to choose the best response according to the degree of risk and proper prevention. The method should be educational and improve its detection ability over time. It should also be designed to maintain the desired level of performance and security with the least computational resources, as the efficiency of cloud services is based on its computing capabilities. Therefore, effective methods should be used to process false-positive signals that support the detection efficiency.

Fig. 4. Block diagram of the MCD in the cloud environment

Method for Cyber Threats Detection and Identification

337

3. Quick detection and warning. Rapid detection and warning is a very important factor in the development of detection methods, as it affects the overall performance of the system and is crucial for the delivery of pre-agreed QoS. A cloud system with multiple administrators should minimize or no human intervention to avoid wasting time on the administrator’s response. It must work in real time and provide automatic responses to suspicious actions. 4. Autonomous self-adaptation. The cyber threat detection system must be selfconfigured and adapted to configuration changes, as computing nodes are dynamically added and removed. The development of an appropriate architecture will allow you to determine how alerts should be processed and distributed from the individual detection components, while maintaining the topological model of cloud computing. It also facilitates the monitoring and control of network components. The design of such a system should be flexible enough to cover future developed standards. 5. Scalability. The MCD must be scalable to efficiently handle the huge number of network nodes available in the cloud and their communication and computational load. 6. Deterministic calculations in the cloud. They provide critical and critical functional services that have specific performance requirements in terms of retention, reliability, and resilience. MCD should not only provide real-time performance, but also ensure the negative impact of the deterministic nature of the network. On the other hand, the performance of monitoring systems should not affect the MCD. The further level of regularity of network traffic within the cloud may change, but the performance of the MCD must remain deterministic. It is necessary to ensure that the MCD has sufficient capacity to process all information. Sharing diagnostic capabilities and computing load on stand-alone agents with the help of the network will increase the level of fault tolerance. 7. Synchronization of autonomous MCD. Information and actions must be synchronized to detect widespread and simultaneous attacks, to apply appropriate responses, or to change a particular component system or configuration of the entire network and to adopt an appropriate prevention strategy. 8. Resistance to compromise. The MCD must protect itself from unauthorized access or attacks. The IMC must be able to authenticate network devices, authenticate the administrator and verify its actions, protect its data, and block any vulnerability that may cause additional vulnerabilities.

4 Experimental Study and Discussion The purpose of the experiments is to test the effectiveness of the developed MCD in cloud services. These experiments will be described in current section. To achieve this goal it is necessary to solve the following tasks: 1) Research of the developed MCD in the developed cloud service (protected data center); 2) Processing and verification of the obtained results; 3) Evaluating the effectiveness of MCD when changing parameters; 4) Determining the possibility of using the developed MCD in other types of IP.

338

Z. Hu et al.

4.1 Experimental Study in RStudio Environment Input/output of the experiment: the input data is 20% of the NSL-KDD data set, the output data are classified data (normal or abnormal – threat). Experimental environment: open-source development environment for R (programming statistics and RStudio data visualization). RStudio includes a console, a syntax highlighting editor that supports direct code execution, and tools for scheduling, logging, debugging, and desktop management. Figure 5 shows the working environment of the RStudio tool.

Fig. 5. Main working window of RStudio tool

The window is divided into four parts: 1) working part – for direct writing and running code, there is also a standard toolbar for all tools; 2) after viewing the data there are tabs of the environment (you can view the loaded data sets and libraries) and history (see versions of the project); 3) console – to display the results of the written program, and data related to the environment (loading of the library); 4) field of view of visualized results (diagrams, histograms, etc.). Stages of research study Stage 1: Connection of all necessary libraries, loading of a training data set of NSLKDD database (Fig. 6–7). Stage 2: Analysis of the test data set Only 20% of the training data of the NSL-KDD database have 25,191 elements that have 43 features (Fig. 8). In Fig. 9 the distribution of types of threats and attacks is shown. As can be seen from Fig. 9 the largest number of attacks is related to the DoS [25]. Stage 3: Direct testing of the method Next, we test our method using built-in functions and data set. Initially, residual data and duplicates were separated, and separate small data sets were identified (Fig. 10).

Method for Cyber Threats Detection and Identification

339

Fig. 6. Code for data set downloading

Fig. 7. Downloaded 20% of training data

After that, the traffic is analyzed (Fig. 11) and the normal and anomalous data are determined, as well as the accuracy of the result (Fig. 12). Stage 4: The result of the experiment During the experiment, the following results were obtained: the total percentage of threats detected – 96.356%, correctly classified – 95.89%, incorrectly classified – 4.11% (Fig. 12). Since the study incorrectly classified a certain percentage of threats, it is logical to apply the definition of errors of the 1st and 2nd kind.

340

Z. Hu et al.

Fig. 8. Attributes of the training data set

Fig. 9. Distribution of types of threats

The first kind of error is that the null hypothesis H0 is rejected, although in reality it is correct. The second kind of error is that the null hypothesis H0 is accepted, although the alternative hypothesis Ha is actually correct. The error of the first type is equivalent to the so-called false positives. An example of an error of the first type. Let’s study a cure for a certain disease. The null hypothesis states that these drugs have no effect on the course of the disease. If we reject the true null hypothesis (we make a type I mistake) and accept the false alternative, that is, we believe that the use of these drugs affects the course of the disease (which, in fact, is not the case). By increasing the confidence level from 95% to 99%, we reduce the probability

Method for Cyber Threats Detection and Identification

Fig. 10. The results of data processing

Fig. 11. The result of traffic analysis

341

342

Z. Hu et al.

Fig. 12. Test results of the MCD

of making a type I error (is rejecting the true null hypothesis) from 5% to 1%. However, there is another danger: this increases the likelihood of making a type II error. Type II error is equivalent to false negatives. The probability of making a type II error is denoted as a. 1 − a - power of the criterion. That is, you need to determine how many of the misidentified threats were actually threats and how many were normal traffic. We have that, IC0 = 5%. Assume that with an improved method algorithm, the percentage of incorrect classifications will decrease. Assume that the value obtained IC = 4.11%, when choosing a larger data set and conducted 10 experiments. The value of the variance is 1%. Solution. Statistics X criterion provided that correct hypothesis H0: IC0 = 4.11% (Fig. 13).

Fig. 13. Calculation of type 2 errors

Result: the probability of receiving errors of 1 type – 5%, 2 types – 12%. 4.2 Experimental Study in the CloudSim Simulation System Input/output of the experiment: the input data is a set of NSL-KDD data and captured network traffic, the output – classified data (normal or abnormal – threat) and the value of the efficiency of the MCD. Experimental environment: CloudSim simulation system.

Method for Cyber Threats Detection and Identification

343

CloudSim platform is a generalized and scalable simulation tool that allows fullfledged modeling and simulation of cloud computing systems and infrastructure, including the construction of data centers using the “cloud”. It is an extension of the basic functionality of the GridSim platform, providing the ability to model data storages, web services, resource allocation between virtual machines [31]. Let’s look at the log of the revealed threats which was written down during carrying out modulation (Fig. 14).

Fig. 14. Log of detected cyber threats

The logs study in RStudio was done to visualize the results: distribution of identified threats and attacks (Fig. 14); diagram of the dependence of the percentage of detection on the type of threat (Fig. 15). As can be seen from Fig. 15 most detected attacks are related to DoS [25] (Fig. 16). Comparison of the results of simulations on the CloudSim platform is shown in Table 2. The displayed results indicate that when simulating the data center model without MCD, but provided that there is a built-in threat level security network detector, the level of detected threats is at 45.87%, which indicates insufficient security of the cloud service, because it means that if only the built-in anti-attack module is present, less than half of the attacks will be detected. And when conducting simulations with the built-in MCD, the level of detected cyber threats is at the level of 93.89%, which indicates the effectiveness of its work. This research study is important because it is directed on practice by using real cyber threats signatures from NSL-KDD data base as well as modern simulation tools RStudio and CloudSim.

344

Z. Hu et al.

Fig. 15. Distribution of identified threats and attacks

Fig. 16. Percentage distribution of identified threats depending on their type

Table 2. Comparison of simulation results Experiment

MCD connection

Detected threats

1

−

45.87%

2

+

93.89%

5 Conclusions and Future Research Study In this paper the analysis of the existing models, systems and MCD was carried out, which allowed to identify their main shortcomings, namely: lack of data on experimental research, the impossibility of its use in cloud services (for the most part), some MCDs do not implement real-time cyber threat detection and others. A model of cloud service has been developed, which uses technological architecture, high-speed communication, unified structures and calculations. It allows to ensure the

Method for Cyber Threats Detection and Identification

345

security of cloud service based on cloud computing and conduct appropriate simulations of cloud service functioning. Improved MCD has been developed, which due to dynamic resources, autonomous self-adaptation and scalability and deterministic calculations allows to detect cyber threats in cloud services and classify them (for example, using to the NSL-KDD classifier or other data bases of cyber threats and classifiers). The developed MCD for cloud services was experimentally investigated using NSL-KDD data base. It has proved the correctness of its work and the possibility of application in cloud services as well as increase efficiency of cloud system security by 48.02% (the efficiency of detecting cyber threats in cloud service is 93.89%, and without the application of the proposed MCD – 45.87%). In addition, a cloud service model has been developed that can be used to build cloud services based on the various cloud computing architecture. In the future, based on the proposed MCD and model, appropriate tools for detecting and classifying cyber threats in cloud services can be developed. It can be autonomous functional unit of SIEM or other instrumental tools of CSIRT/SOC.

References 1. Abidar, R., Moummadi, K., Moutaouakkil, F., Medromi, H.: Intelligent and pervasive supervising platform for information system security based on multi-agent systems. Int. Rev. Comput. Softw. 10(1), 44–51 (2015) 2. Ivanov, A.: Security as main pain of the cloud computing, Online access mode. http://www. cnews.ru/reviews/free/saas/articles/articles12.shtml 3. Active security for advanced threats counteraction, Online access mode. http://www.itsec.ru/ articles2/target/aktivnaya-zaschita-kak-metod-protivodeystviya-prodvinutym-kiberugrozam 4. The 6 Major Cyber Security Risks to Cloud Computing, Online access mode. http://www. adotas.com/2017/08/the-6-major-cyber-security-risks-to-cloud-computing/ 5. Google Security Whitepaper for Google Cloud Platform, Online access mode. https://habrah abr.ru/post/183168/ 6. Dokas, P., Ertoz, L., Kumar, V.: Data mining for network intrusion detection. Recent Adv. Intrusion Detect. 15(78), 21–30 (2014) 7. Ahmed, P.: An intrusion detection and prevention system in cloud computing: a systematic review. J. Netw. Comput. Appl. 11, 1–18 (2016) 8. Anderson, J.P.: Computer security threat monitoring and surveillance. Tech. Rep. Contract 36, 179–185 (1982) 9. Carl, G., Kesidis, G., Brooks, R.R., Rai, S.: Denial-of-service attack-detection techniques. Internet Comput. IEEE 10, 82–89 (2006) 10. Hu, Z., Gnatyuk, V., Sydorenko, V., et al.: Method for cyberincidents network-centric monitoring in critical information infrastructure”. Int. J. Comput. Netw. Inf. Secur. 9(6), 30–43 (2017). https://doi.org/10.5815/ijcnis.2017.06.04 11. Chatzigiannakis, V., Androulidakis, G., Maglaris, B.: A Distributed Intrusion Detection Prototype Using Security Agents, HP OpenView University Association, pp. 14–25 (2004) 12. Abraham, T.: IDDM: intrusion detection using data mining techniques. DSTO Electron. Surveill. Res. Lab. 9, 30–39 (2001) 13. Zaliskyi, M., Odarchenko, R., Gnatyuk, S., Petrova, Y., Chaplits, A.: Method of traffic monitoring for DDoS attacks detection in e-health systems and networks. In: CEUR Workshop Proceedings, 20186, vol. 2255, pp. 193–204 (2018)

346

Z. Hu et al.

14. Chouhan, M.: Adaptive detection technique for cache-based side channel attack using bloom filter for secure cloud. Conf. Comput. Inf. Sci. 1, 293–297 (2016) 15. Zhengbing, H., Gnatyuk, S., Koval, O., Gnatyuk, V., Bondarovets, S.: Anomaly detection system in secure cloud computing environment. Int. J. Comput. Netw. Inf. Secur. 9(4), 10–21 (2017). https://doi.org/10.5815/ijcnis.2017.04.02 16. Li, H.-H., Wu, C.-L.: Study of network access control system featuring collaboratively interacting network security components. Int. Rev. Comput. Softw. 8(2), 527–532 (2013) 17. Dilek, S., Çakır, H., Aydın, M.: Applications of artificial intelligence techniques to combating cyber crimes: a review. Int. J. Artif. Intell. Appl. 6(1), 21–39 (2015) 18. How Big Data Can Improve Cyber Security, Online access mode. https://csce.ucmss.com/cr/ books/2017/LFS/CSREA2017/ABD3239.pdf 19. Kirichenko, L.: Cyber threats detection using social networks analysis. Int. J. Inf. Technol. Knowl. 11, 23–32 (2017) 20. Cisco creates self-defending networks for cyber threats detection, Online access mode. https://nag.ru/news/newsline/30762/v-cisco-sozdayut-samooboronyayuschuyusya-set-dlyavyiyavleniya-kiberugroz.html 21. Xiaohua, Y.: Early detection of cyber security threats using structured behavior modeling. ACM Trans. Inf. Syst. Secur. 5, 10–35 (2013) 22. Methods for deep analytics to counteract of modern threats, Online access mode. http://bisexpert.ru/sites/default/files/archives/2016/bis9_konovalov.pdf 23. Sakr, M.M., Tawfeeq, M.A., El-Sisi, A.B.: An efficiency optimization for network intrusion detection system. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 11(10), 1–11 (2019). https://doi. org/10.5815/ijcnis.2019.10.01 24. Pat. No WO2015159287. System and method for cyber threats detection; author Malachi Y.; Accessed 22 October 2015 25. Hassan, Z., Odarchenko, R., Gnatyuk, S., et al.: Detection of distributed denial of service attacks using snort rules in cloud computing & remote control systems. In: Proceedings of the 2018 IEEE 5th Intern. Conf. on Methods and Systems of Navigation and Motion Control, 16–18 October 2018. Kyiv, Ukraine, pp. 283–288 (2018) 26. Byrski, A., Carvalho, M.: Agent-based immunological intrusion detection system for mobile ad-hoc networks. In: Bubak, M., van Albada, G.D., Dongarra, J., Sloot, P.M.A. (eds.) ICCS 2008. LNCS, vol. 5103, pp. 584–593. Springer, Heidelberg (2008). https://doi.org/10.1007/ 978-3-540-69389-5_66 27. Zhang, Z.: Hide: a hierarchical network intrusion detection system using statistical preprocessing and neural network classification. IEEE Workshop Inf. Assur. Secur. 16, 85–90 (2001) 28. Pat. No US20020038430 A1. System and method of data collection, processing, analysis, and annotation for monitoring cyber-threats and the notification thereof to subscribers; authors Charles Ed., Samuel M., Roger N., Daniel O.; Accessed 23 March 2012 29. Pat. No US9749343B2. System and method of cyber threat structure mapping and application to cyber threat mitigation; authors John P, Frederick D., Henry P., Keane M.; Accessed 4 March 2013 30. Arora, I.S., Bhatia, G.K., Singh, A.P.: Comparative analysis of classification algorithms on KDD’99 data set. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 8(9), 34–40 (2016). https://doi. org/10.5815/ijcnis.2016.09.05 31. Buyya, R., Ranjan, R., Calheiros, R.: Modeling and simulation of scalable cloud computing environments and the CloudSim toolkit: Challenges and opportunities. In: International Conference on High Performance Computing Simulation USA, IEEE, pp. 1–11 (2009)

Nonparametric Change Point Detection Algorithms in the Monitoring Data Igor Prokopenko(B) National Aviation University, Kyiv, Ukraine

Abstract. Monitoring systems are used for a lot of applications: environment monitoring, reliability monitoring, traffic monitoring in telecommunication systems etc. An important task in all systems is to identify the moment of change in the state of the observed process, and to make operational decisions on its correction. Characteristics of the effectiveness of decision-making methods are errors of the first and second kind - the probability of false alarms and the probability of omission. The method of cumulative sums CUSUM is most often used to identify the moment of change in the statistical properties of the process. Usually technical systems work in difficult conditions (change of temperature, external loading, change of noise situations as a result of counteraction that leads to need of adjustment or adaptation of parameters of algorithm, in particular, a decision-making threshold. In such conditions it is expedient to apply the nonparametric approach to synthesis of data processing algorithms. Nonparametric algorithms ensure the stability of the probability of false alarm at the intervals of process stationarity. The theory of synthesis of nonparametric rank algorithms for detecting changes in the statistical characteristics of the process is presented and their efficiency is investigated. Keywords: Monitoring systems · Change point detection · Rank nonparametric algorithms · Telecommunication system traffic measurement

1 Introduction Monitoring the state of information systems for various purposes, such as computer networks, creates the task of detecting changes in the statistical properties of the studied process (traffic in computer networks, electricity quality in energy management systems, etc.). Statistical properties can vary widely. The parameters of known probability distributions can change - shear parameters, scale parameters. Changing the mode of operation of the system (network congestion, attack, emergency) leads to a change in the law of distribution of process probabilities. Nonparametric decision rules must be applied to detect changes in the statistical characteristics of the process.

2 Related Papers Analysis The onset of nonparametric detection theory was recorded in 1936 when Hoteling and Pabst published an article on rank correlation. Since then, a large number of papers © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 347–360, 2021. https://doi.org/10.1007/978-3-030-80472-5_29

348

I. Prokopenko

on nonparametric and robust methods of information processing have been published [1–12]. Although nonparametric or distribution-free methods have been studied and applied by statisticians for at least 50 years, only recently have engineers recognized their importance and begun to investigate their applications. In particular, in recent years, much attention has been paid to the study of methods for detecting traffic anomalies in computer networks caused by DDoS attacks [17–21]. In these problems, an important characteristic is the stability of the probability of false detection of the attack, which can be ensured by the application of non-parametric decision-making rules. In this paper, we consider the problem of synthesis of rules of nonparametric solutions to identify changes in the statistical characteristics of the controlled process. The structure of the article includes the formulation of the problem of detecting a disorder of a random process and determining the point of change of its characteristics (i.3), a description of methods for constructing nonparametric, free from the distribution, decisive rules for testing statistical hypotheses about the stationarity of a random process and modifications of two-sided rank nonparametric Wilcoxon algorithm (i.4), that use to detect the start time and stop attacks on computer networks. A feature of this problem is Poisson’s law of traffic distribution. In item 5) and item 6) by the Monte Carlo method the efficiency of modifications of algorithm in a problem of change of statistical characteristics of Poisson’s random process is investigated, influence of deviations from Poisson’s law is estimated. In item 6.3, a new optimal two-sample rank algorithm for detecting changes in the characteristics of a random process with a “contaminated” Poisson distribution is synthesized, expressions for rank probabilities are obtained, and a comparative analysis of the efficiency of nonparametric rank algorithms and CUSUM algorithm in the computer network problem.

3 Statement of the Problem Consider a time series formed by time-discrete measurements of a random process x1 , . . . .xn .

(1)

Let the statistical characteristics of the process change at the k-th moment of time. That is, by the time t1 , . . . tk−1 the process has a multidimensional probability distribution ω0 (x1 , .., xk−1 ),

(2)

and after the k-th moment multidimensional probability distribution is ω1 (xk , .., xn ).

(3)

The task is to identify the fact of change in statistical characteristics and assess the moment of change. This problem in statistics is called the change point detection problem.

Nonparametric Change Point Detection Algorithms

349

4 Nonparametric Methods for Solving Change Point Detection Problem To construction of free of distribution procedures for the transferred verification tasks of nonparametric hypotheses are devoted the next parts of this chapter. 4.1 Statisticians, Which are Used in Nonparametric Tasks Construction of a statistics, free of distribution, is based on that fact, that the random variables ri , got by transformation ______

ri = Fy (xi ), i = 1, k − 1

(4)

where xi are values of random variable X, F(.) is cumulative distribution function of random variable X, have the uniform distribution on an interval (0,1). In nonparametric tasks there is cumulative distribution function F (x) unknown. Therefore, in accordance with empiric Bayesian approach, one can use the training sample xk , .., xn to get estimation of distribution function, and use this estimation for construction of nonlinear transformation (4). Thus, samples x1 …xk-1 are transformed in accordance with an algorithm ______

ri = Fy∗ (xi ), i = 1, k − 1

(5)

and a hypothesis H0 about uniform distribution of a random variable R is checked. If this hypothesis is adopted, we make conclusion, that a hypothesis H0 takes place about the samples xk , .., xk−1 and xk , .., xn belong to the same distribution, that is the sample x1 , .., xn does not contain change point. In the opposite case makes the decision about the presence of change point in a sample x1 , .., xn . In nonparametric tasks, when about the form of the output distribution nothing is known, the set of order statistics will be minimum sufficient statistics of the sample xk , .., xn or variation series y(1) = min(xk , ..xn ) < y(2) < . . . < y(m) = max(xk , ..xn ), m = n − k + 1. A variational series is used for construction of empiric function of probability distribution Fx∗ (x) =

______ i , y(i) < x ≤ y(i+1) , i = 1, m m+1

(6)

A function (6) is a step function with the breaks of the first type in points, which are equal to the variable of order statistics. It is used as estimation of cumulative distribution function of random variable X – F(x).

350

I. Prokopenko

4.2 Methods of Synthesis of Free Distribution Algorithms of Signals Detection At construction of algorithm of signal detection it is possible to go out from the _____ task of verification of hypothesis H0 about evenness of distributing of variables r i , i = 1, k − 1, which are determined by transformation (5). These variables correspond to the ranks of _____ samples x i , i = 1, k − 1 concerning variation series y(1) , ..y(m) . Really, let’s consider the joined sample xi , y(1) , ..y(m) , build a variation series and define the place of x i out in it that is calculate its rank – Ri . Let Ri = k, where k – integer from an interval [1, m + 1]. It means, in accordance with (6), that y(k−1) < xi < y(k) ri =

Ri k = m+1 m+1

(7)

A formula (7) gives relation between the rank of the of x i out in the variation series formed from the variation series y(1) , ..y(m) and sample x i , and value of function F ∗ (xi ) F ∗ (xi ) =

Ri m+1

(8)

One of the decision rules synthesis methods of hypothesis H0 verification is based on the use of concordance criteria of the empiric distribution of random variable ri , i = 1, k − 1, with the discretely uniform distribution on an interval [0,1]. 4.3 Wilcoxon’s Criterion Two-sample criterion is offered by Wilcoxon [12]. This criterion calculates the sum of ranks of sample x 1 …x n concerning to the variation series y(1) , ..y(m) . S=

n 

Ri

(9)

i=1

where the rank of sample x i relative the variation series y(1) , ..y(m) is calculated by formula Ri =

m  j

sgn(z) =

sgn(xi − yj ), i = 1, n; 

1, z > 0, 0, z 3, distribution of statistics (9) by H0 can be assumed as Gaussian with parameters nm , 2 nm μ2 {S|H0 } = (n + m + 1). 12 m1 {S|H0 } =

(11)

The rule of verification of hypothesis H0 about belonging of sample to the same distributing after a criterion of Wilcoxon is formulated as follows.

Nonparametric Change Point Detection Algorithms

351

At the one-sided alternative H+ 1:

F1x (z) > F1y (z)

(12)

hypothesis H0 is rejected, if at the given level the probability of false alarm is  (13) S − m1 {S|H0 } > Vp = χ1−α μ2 {S|H0 }, where χ1−α - quantile of 1-α level of the normalized Gaussian distribution. Consider Signal detector that realizes a two-sided criterion of Wilcoxon. At a two-sided alternative H1 : F1x (z) = F1y (z) one can use the following statistic Z = |S − m1 {S|H0 }|.

(14)

Variance of this statistic is μ2 {Z|H0 } = 0.345

nm (n + m + 1), 12

(15)

and mean value  nm m1 {Z|H0 } = 0.8 (n + m + 1). 12

(16)

Equations (15) and (16) were obtained by numerical experiment and allows to calculate mathematical expectation and variance of statistic (14) with accuracy 0.5%–1.5%. A hypothesis H0 reject, if at required probabilities of false alarm α inequality is executed  (17) Z = |S − m1 {S|H0 }| > Vd = m1 {z|H0 } + χ1−α μ2 {Z|H0 }, where χ1−α - quantile of 1 − α level of normalized Gaussian distribution. An one-sided algorithm of Wilcoxon differs from two-sided by absence of blocks of calculation of difference and calculation of the module. Taking into account expression for the calculation of rank (10) realization of Wilcoxon’s algorithm can be simplified. It is possible to give up the blocks of organization of selections (index filters), replaces them by the blocks of memory (BM) of selections, and to realize the calculators of ranks according to (10). Wilcoxon’s algorithm provides independence of probability of false alarm from the form of law of random process distribution, but a different efficiency is had at different alternatives. An algorithm of Wilcoxon provides greater efficiency in tasks, when the shift parameter of the process is changed.

352

I. Prokopenko

5 Application of Non-parametric Wilcoxon Algorithm for Detection of Start and Termination Time of the Attacks on Computer Networks Network traffic contains large amounts of data and requires ongoing monitoring. The tasks of traffic monitoring are the analysis of the current state of the network and the detection of anomalies caused by changes in the state of the network or DoS attack. Statistics of the operation of computer networks show an increasing number of DoS attacks, which leads to a violation of the integrity of information and losses of companies. Therefore, statistical analysis of network traffic and the search for effective algorithms for detecting abnormal behavior of network traffic and the inclusion of countermeasures is an urgent task. In [13], the classification of network attacks is considered and the application of the algorithm of cumulative sums CUSUM [14–17] is proposed to identify the moment of traffic disorder that occurs in the event of an attack. Then its intensity sharply increases. This algorithm, along with certain advantages, has disadvantages. It does not provide stability for the probability of false alarms in different modes of network operation. Consider the problem of detecting a disorder of computer network traffic in nonparametric and setting according to paragraph 2. The current monitoring will be carried out in a sliding window of 2m + 1. In the center of the window is the point of the process, which corresponds to the k-th moment of decision-making. On the first half of the window we will form a sample x1 , . . . .xm , on the second half of the window we will form a sample y1 , . . . .ym . At the m + 1 point we will test the nonparametric hypothesis H0 : ω0 (x1 , .., xm ) = ω1 (y1 , .., ym ) against the alternative H1 : ω0 (x1 , .., xm ) = ω1 (y1 , .., ym ) We apply the two-way Wilcoxon algorithm (17–20). Z = |S − m1 {S|H0 }| > Vd , where the mathematical expectation and variance of the statistics Z are estimated by formulas (15–16) at n = m.  m2 m1 {Z|H0 } = 0.8 (2m + 1), 12 m2 (2m + 1). 12 Decision threshold according to formula (17) is  Vd = m1 {Z|H0 } + χ1−α/2 μ2 {Z|H0 }. μ2 {Z|H0 } = 0.345

Quantiles of normalized Gaussian distribution for different levels are given in the Table 1. The structure of two-side attack detection algorithm is show on Fig. 1 (Fig. 2).

Nonparametric Change Point Detection Algorithms

353

Table 1. Quantiles of normalized Gaussian distribution α

0.1

0.01

0.005

0.001

0.0005

0.0001

0.00001

0.000001

χ1−α/2

1.64

2.58

2.8

3.29

3.55

3.9

4.42

4.89

Fig. 1. Two-side rank detector of DoS attack

Fig. 2. One-side rank detector of DoS attack

6 Properties Analysis The Monte Carlo method was used to analyze the efficiency of the algorithms. This method consists in modeling a random process with given characteristics, modeling the algorithm for processing the sequence of its values and calculating the characteristics of the efficiency of the algorithm. For detection algorithms, the dependence of the probability of detecting an attack on its intensity, the probability of false alarm and the accuracy of estimating the moments of the beginning and end of the attack are estimated.

354

I. Prokopenko

6.1 Study of Algorithms Efficiency in the Distribution of Network Traffic According to Poisson’s Law Poisson’s flow is used as the main model of network traffic. We enhance it in the following way to take into account difference from Poisson distribution. The distribution of the number of packets per unit time is given by the expression f (x, t) = (1 − p)

λ(t)x exp(−λ(t)), x = 0, 1..∞, x!

where a random variable takes values from a set of integers; λ(t) - flow intensity, which can be constant in the case of a stationary flow, or change over time. DoS attacks are characterized by a sharp increase in the intensity of the flow during the attack. This feature is used to detect them. To evaluate the effectiveness of the algorithm for detecting and evaluating the parameters of the DoS attack, a sequence of values of the Poisson process with variable intensity is odeled.  λ1 , t ∈ [t1 , t2 ]; λ(t) = / [t1 , t2 ]. λ0 , t ∈ The detection algorithm captures the fact of the attack and evaluates its beginning and end, ie the values t1∗ and t2∗ . Simulation results are shown in Fig. 3. The average graph shows the implementation of the Poisson process. In stationary mode, the flow rate is 4 packets per second. DoS attack continued with t1 = 100 sec . to t2 = 160 sec . During the attack, the intensity increased to 9 packets per second. The upper graph shows the output of the threshold device, which forms one in case of exceeding the decision threshold and zero, if the threshold is not exceeded. Single pulses are visible, signaling the beginning and end of the attack around the points t1 = 100 sec . i t2 = 160 sec . In the vicinity of the points t3 = 405 sec ., t4 = 425 sec . and t5 = 492 sec . false signals are observed. The following graph shows the values of the test statistics of algorithm. The position of peaks gives estimation of the moments of attack start and finish t1∗ and t2∗ . Figure 4 shows the position of the peaks on an enlarged scale. It can be seen that the beginning and end of the attack are recorded to the nearest second. In Fig. 5 shows the characteristics of attack detection for algorithm (14) (Fig. 7). Wilcoxon’s one-side algorithm provides a higher probability of detecting an attack with the same sample size n, but it detects only the beginning of the attack, as shown in Fig. 6. 6.2 Research of Algorithms Robustness To investigate robustness of suggested algorithms we enhance Poisson’s model in the following way to take into account difference from Poisson distribution.

Nonparametric Change Point Detection Algorithms

355

Fig. 3. The results of modeling the two-side Wilcoxon algorithm (14).

Fig. 4. The position of the detector peaks on an enlarged scale

The distribution of the number of packets per unit time is given by the expression f (x, t) = (1 − p)

p x λ(t)x exp(−λ(t)) + L exp(− ), x = 0, 1..∞, x! L L

(18)

where a random variable takes values from a set of integers; λ(t) - flow intensity, which can be constant in the case of a stationary flow, or change over time; L-scale parameter; p-probability of jammer action. We investigate efficiency and robustness of four change point detection algorithms: cumulative sum CUSUM (

2m 

i=m+1

xi −

m  i=1

xi ) > Vd ,

(19)

356

I. Prokopenko

Fig. 5. Detection characteristics of of two-side Wilcoxon detector (14)

Fig. 6. The results of modeling one-side Wilcoxon algorithm (13)

Wilcoxon’s rank algorithm (13) and optimized rank algorithm. 6.3 Optimization of Rank Detection Algorithm Optimization of ranking criteria can be done by calculating the probabilities of ranks for an alternative hypothesis H1 . Let the alternative hypothesis be that when H1 the intensity

Nonparametric Change Point Detection Algorithms

357

Fig. 7. Detection characteristics of one-side Wilcoxon detector (13)

of the flow increases and becomes. λ1 = λ0 + b. Then probabilities of ranks H1 can be calculated according to the formulas (19)   m m   fp (k, λ + b)FpRi (k, λ)(1−Fp (k, λ))m−Ri , i = 1, m P1 (Ri ) = Ri k=0





where fp (∗) − Poisson s PDF; FpRi − Poisson s CDF

(20)

and the statistics of the test criterion H1 against H0 , according to the NeumannPearson lemma, takes the form 2m 

log P1 (Ri ) > Vd

(21)

i=m+1

We investigate influence of deviation from Poisson’s model by changing of parameter p in (18) and correlation coefficient of neighboring samples ρ{xi , xi+1 } on algorithm properties by the transformation  (22) xi+1 = [ρxi + 1 − ρ 2 ςi+1 ], Where ςi+1 -Poisson’s random number, [*] - operation of taking the whole part (Figs. 8, 9). Analysis of obtained results shows strong influence of deviation from Poisson’s distribution law on efficiency of CUSUM algorithm and robustness of optimal rank and Wilcoxon’s algorithms. Correlation of neighboring samples led to strong decreasing of efficiency of CUSUM algorithm and to small decreasing of rank algorithms efficiency. But optimal rank algorithm is more robust then Wilcoxon’s algorithm.

358

I. Prokopenko

Fig. 8. Operating characteristics of the change point detectors (13), (19), (21). n = m = 10; probability of false alarm F = 0.01; p = {0;0.01}; non correlated samples (correlated coefficient ρ = 0)

Fig. 9. Operating characteristics of the change point detectors (13), (19), (21). n = m = 10; probability of false alarm F = 0.01; p = {0;0.01}; correlated samples (correlated coefficient ρ = 0.2)

7 Conclusion Studies have shown that nonparametric methods for detecting nonstationarity provide stability to the probability of false alarm when changing the parameters and shape of the probability distribution law of the controlled process. Wilcoxon’s ranking algorithm provides a fairly high efficiency of detecting changes in the Poisson flux intensity parameter, losing in the threshold signal at the level of D = 0.5 to the optimal CUSUM algorithm at about 4.5 dB. However, CUSUM does not provide stability of the probability of false alarm when changing the parameters of the scale and shape of the probability distribution in stationary mode. The efficiency of a nonparametric ranking algorithm can be increased by calculating the probabilities of ranks under the alternative hypothesis. The reduction of the threshold signal at the level of D = 0.5 can reach from 1.5 to 2 dB. Application of nonparametric rank algorithms provides stability of false alarm probability and allows us to increase defense ability of computer networks.

Nonparametric Change Point Detection Algorithms

359

References 1. Huber, P., Ronchetti, E.: Robust Statistics. 2nd Edition, A John Wiley and Sons, Inc., Publication, p. 380 (2009). https://doi.org/10.1002/9780470434697 2. Hampel, F.: A general qualitative definition of robustness. Ann. Math. Stat. 42, 1887–1896 (1971). https://doi.org/10.1214/aoms/1177693054 3. Hogg, R.: Introduction to noise-immune estimation. In: Stable statistical methods of data evaluation. Moscow, Mechanical Engineering (1984) 4. Kornilev, E., Prokopenko, I., Chuprin, V.: Ustoychivye algoritmy v avtomatizirovannyh sistemah obrabotki informacii. Technica, Kyiv (1989). (Russian) 5. Solomentsev, O., Zaliskyi, M., Kozhokhina, O., Herasymenko, T.: Reliability parameters estimation for radioelectronic equipment in case of change-point. In: Signal Processing Symposium (SPS-2017), 12–14 September 2017, (Jachranka Village, Poland), Proceedings, pp. 1–4 (2017) 6. Rohling, H.: Radar CFAR thresholding in clutter and multiple target situations. IEEE Trans. Aerosp. Electron. Syst. Aes-19(4), 608–621 (1983). https://doi.org/10.1109/TAES.1983. 309350 7. Prokopenko, I.: Robust methods and algorithms of signal processing. In: IEEE Microwaves, Radar and Remote Sensing Symposium (MRRS-2017), 29–31 August 2017, Kyiv, Ukraine, Proceedings, pp. 71–74 (2017). https://doi.org/10.1109/MRRS.2017.8075029 8. Prokopenko, I.G.: Statistical synthesis of robust signal detection algorithms under conditions of aprioristic uncertainty. Cybern. Inf. Technol. 15(7), 13–22 (2015). https://doi.org/10.1515/ cait-2015-0085 9. Hampel, F.: Robust estimation: a condensed partial survey. Z. Wahr. Verw. Geb. 27, 87–104 (1973). https://doi.org/10.1007/BF00536619H 10. Kassam, S., Poor, H.: Robust signal processing for communication systems. IEEE Commun. Mag. 21, 20–28 (1983). https://doi.org/10.1109/MCOM.1983.1091322 11. Hajek, J., Sidak, Z.: Theory of Rank Tests, 2rd ed., Academic Press (1999). https://doi.org/ 10.1016/B978-012642350-1/50021-7 12. Kendall, M., Stuart, A.: Statistical inference and connections. Inference and Relationship, Published by International Statistical Institute, vol. II, 1976, pp. 167–169 (1976) 13. Addie, R.G., Neame, T.D., Zukerman, M.: Performance evaluation of a queue fed by a Poisson Pareto burst process. Comput. Netw. 40(3), 377–397 (2002). https://doi.org/10.1016/S13891286(02)00301-8 14. Alippi, C., Roveri, M.: An adaptive cusum-based test for signal change detection. In: Proceedings - IEEE International Symposium on Circuits and Systems, January 2006 15. Feinstein, L., Schnackenberg, D., Balupari, R., Kindred, D.: Statistical approaches to DDoS attack detection and response. In: Proceedings DARPA Information Survivability Conference and Exposition (2003). https://doi.org/10.1109/discex.2003.1194894 16. Zalisky, M., Odarchenko, R., Gnatyuk, S., Petrova, Y., Chaplits, A.: Method of traffic monitoring for DDoS attacks detection in e-health systems and networks. In: CEUR Workshop Proceedings, vol. 2255, pp. 193–204 (2018) 17. Hu, Z., et al.: Statistical Techniques for Detecting Cyberattacks on Computer Networks Based on an Analysis of Abnormal Traffic Behavior. Int. J. Comput. Netw. Inf. Secur. (IJCNIS), 12(6), 1–13 (2020). https://doi.org/10.5815/ijcnis.2020.06.01 18. Kamel, S.O.M., Elhamayed, S.A.: Mitigating the impact of IoT routing attacks on power consumption in IoT healthcare environment using convolutional neural network. Int. J. Comput. Netw. Inf. Secur. (IJCNIS) 12(4), 11–29 (2020). https://doi.org/10.5815/ijcnis.2020.04.02

360

I. Prokopenko

19. Süzen, A.A.: A risk-assessment of cyber attacks and defense strategies in industry 4.0 ecosystem. I. J. Comput. Netw. Inf. Secur. 12, 1–12 (2020) 20. Kim, H., Rozovskii, B.L., Tartakovsky, A.G.: A nonparametric multichart cusum test for rapid detection of DoS attacks in computer networks. Int. J. Comput. Inf. Sci. 2(3), 149–158 (2004). On-Line 21. Hamolia, V., Melnyk, V., Zhezhnych, P., Shilinh, A.: Intrusion detection in computer networks using latent space representation and machine learning. Int. J. Comput. 19(3), 442–448 (2020)

Computer Science for Education, Medicine and Biology

Combinatorics of the Ford-Fulkerson Algorithm to Reduce the Risks of the COVID-19 Pandemic in International Tourism Marharyta Sharko1(B) , Vira Fomishyna1 , Olha Liubchuk2 , Natalia Petrushenko1 , Tetiana Yakymchuk1 , and Liliia Chaika-Petehyrych1 1 Kherson National Technical University, Kherson, Ukraine

[email protected] 2 State Higher Educational Institution, “Pryazovskyi State Technical University”,

Mariupol, Ukraine [email protected]

Abstract. This paper offers a solution of the urgent problem of modern world tourism management conditioned by the risk and uncertainty of the external environment influence on the international tourism services consumption. The combinatorics of international travel services realization is based on the Ford-Fulkerson algorithm. The novelty of the methodology consists in the transformation of services at the time of the introduction or removal of coronavirus-bound pandemic restrictions. At the same time, the distribution of risks on the range of modified tourist services, namely the risks from possible termination of tourist travel, as well as those from a non-disposal of the offered tourist products, is achieved. The combination of the tourist services disposal at different stages of the trip in the form of a directed graph of tourist services as a part of a tourist product is presented. The main incentive for the implementation of the offered methodology is to maintain the income of manufacturers and suppliers of tourism services and to satisfy consumers by expanding possible offers. The use of combinatorics of the offered services in the state space, their implementation at different moments of iterations, their application with the synchronization of the financial flows throughput allows to reduce risks in the tourist product acquisition and its further disposal. The value of the work lies in reducing the risks of international tourism in the context of the COVID-19 pandemic and preserving the income of manufacturers and suppliers of tourism services. Keywords: Transformation · Combinatorics · Travel services · Risk distribution

1 Introduction In the face of global risks associated with limited movement and self-isolation due to the coronavirus, the global tourism industry is in an extremely difficult situation. It becomes impossible to predict the development of the situation even in the short term. Despite a number of preventive measures and restrictive measures, the tourism industry © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 363–373, 2021. https://doi.org/10.1007/978-3-030-80472-5_30

364

M. Sharko et al.

is developing under conditions of uncertainty and unpredictability of the influences of environmental factors. Tourism management in conditions of dynamic changes in the external environment does not have a developed methodological base. Management decisions are made on the basis of stochastic recommendations based on the results of existing experience with extrapolation to future trends without taking into account risks and possible errors. This introduces a great deal of uncertainty and inadequacy in decision-making related to the multicriteria of assessments, a multitude of dimensions of factors, and multidirectional indicators. The coronavirus has sharply redrawn the world tourism map. Developing rapidly, it acquires the character of an international calamity. According to the World Business Travel Association, the tourism industry has lost up to 300 million jobs. Air transportation decreased by 20%, the demand for travel fell by 25% [1]. The geography of political and natural risks is expanding [2]. In this situation, neither the consumer of a tourist product, nor its manufacturer can predict the consequences of choosing the appropriate alternative. Manufacturers of tourism services suffer heavy losses because of the termination of their implementation at any stage of implementation. The stereotypes of management and marketing as part of strategic management require updating. The aim of the paper is to develop a methodology for managing the development of world tourism enterprises in the context of COVID-19 based on the use of the connection between management theory and methods of transforming complex structures so that the negative influence of external excitations reaches a minimum in situations adequate to external manifestations.

2 Relative Works The efficiency of enterprises management in the field of international tourism is based on constant innovations, changes in their assortment, offers, prices, terms of sale [3–5]. To reduce the level of risk and the results of possible damage associated with the uncertainty of the implementation of tourism services, one should not rely entirely on the organizational efforts of government regulators [6, 7]. It is necessary to examine carefully all possible risk carriers, taking into account their individual characteristics, as well as market participants with the development of their own original risk management methods [8–10]. The initial information on identifying the problems of unstable development of the tourism market is contained in the ratio of internal and external destabilizing factors [11, 12]. The problem is that the subjects of the tourism market cannot influence external factors, but they must be taken into account [13–15]. Information, as an integral part of doing business, plays a decisive role in reducing the risks that ensure the commercialization of offers. Marketing of information products and services is manifested in the ability to study the market environment using information flows and sources of its receipt in the context of new technological opportunities [16–18]. In fact, only online businesses can sustain tourism in the COVID-19 pandemic. Graph theory using the Ford-Fulkerson algorithm is one of the promising areas of analysis and development management under conditions of uncertainty and ambiguity

Combinatorics of the Ford-Fulkerson Algorithm

365

of the external environment [19–21]. Recently, this trend of research has been gaining the greatest popularity. A comprehensive literature review shows that the Ford-Fulkerson algorithm is used to find the maximum flow in the transport network [22–24]. In modern scientific and technical literature there are a large number of publications on the implementation of this algorithm for solving various technical problems. So, in [11] it was applied to solve combinatorial structures of weighted graphs with a low field volume, in [25] when optimizing train routes to control the throughput of rolling stock, in [21, 26] for analyzing bottlenecks in traffic, in [17] for finding the maximum flow with the smallest number of iterations, in [12, 27] for calculating the distance in metal structures under loading. In [28], the application of the Ford-Fulkerson algorithm to identify redundant information is considered. The idea of the algorithm was developed in the combinatorics of the Ford-Fulkerson algorithm to reduce the risks of the covid-19 pandemic in international tourism. The algorithm has been improved many times as a means of maintaining the maximum flow in an extensive transport network [29–31].

3 Materials and Methods Changing conditions for the functioning of tourism enterprises of international tourism, especially in case of restrictions caused by COVID-19, are forcing the transition to fundamentally new models of management and forecasting based on the implementation of iterations of stepwise use and realization of opportunities with a forced transition to other ways of interaction between the supply and consumption of international tourism services. Expert assessments of the attractiveness of various types of tourist services, experience of their use, priority of consumption and the possibility of reducing risks from unpredictable influences of the external environment are used as the material of the study [32]. The methods of preserving the capacity of the transport linear programming problem, which are solved using the Ford-Fulkerson algorithm, the construction of a residual network and combinatorics of branches of tourist services offers with the transition of iterations to different time intervals of the implementation of a tourist product are used [25, 33]. The formal description of Ford-Fulkerson algorithm is as follows: given a graph G (V, E) with capacity c (u, v) and flow f (u, v) = 0 for edges from u to v. It is necessary to find the maximum flow from source s to drain t. At each step of the algorithm, identical conditions are applied for all flows: f (u, v) ≤ c (u, v). i.e. the flow from u to v does not exceed the capacity. f(u, v) = f(u, v). 

f (u, v) = 0 ← → fin (u) = fout (u) for all nodes u except s to t i.e. the flow does not change as it passes through the node. The flow in the network is equal to the sum of the flows of all arcs incident to the flow of the graph.

366

M. Sharko et al.

Marking is used to find a chain along which a gradually increasing flow can be sent. Each arc is assigned a weight, which is written as a fraction on the edge of the original network.

4 Methodology Risk administration in conditions of uncertainty means the need to choose a solution option, when an alternative decision-making has many outcomes, the probabilities of which are unknown. In this situation, the COVID-19 pandemic carries not only inevitable losses, but also the probability of new development. The methodology for studying its dynamics can be disseminated using graph theory and the Ford-Fulkerson algorithm. The paper uses the methodology of maintaining the size of financial flows through the stepwise use and implementation of opportunities with a forced transition to other ways of interaction between the supply and consumption of international tourism services. According to the Ford-Fulkerson algorithm, the network is considered as a connected digraph oriented in one direction. The algorithm makes it possible to improve the linear programming problem. It includes stream persistence properties and is pseudopolynomial, i.e. has an estimate: 0(nmlogu), where m = |E|, n = |V |, u = max Cij

(1)

The algorithm starts from zero flow and increases the network flow at each iteration until the network arc becomes saturated. The flow in the network, for which it is impossible to build an increasing chain, is the maximum. When building a network from the source to the drain, placement of labels is used in order to determine the amount of flow by which its value can be changed.

5 Experiment Specifying of application of the Ford-Fulkerson algorithm at various iterations of decision making is associated with combinatorics. Methodological support for the preservation of financial flows of external tourism in the context of global risks is associated with the synchronization of its use, depending on consumer preferences and forced measures to strengthen or limit the offers. The analysis of statistical reporting prior to COVID-19 made an opportunity to determine the weights of the digraph arcs of the tourist network through experimental evaluation, which are presented in Table. 1. The analysis was carried out on the basis of processing the results of the “Tour PLAZA” travel agency. Travel duration in days is selected as network nodes. Usually it lasts a week, so the number of nodes is chosen equal to 7. The network of the tourist product sales is shown in Fig. 1. The occupancy of the days of stay with the possibility of participating in the services offered is presented in columns 1–4, 5–8, 9–12, 13–16, 17–20, 21–24, 25–28.

Combinatorics of the Ford-Fulkerson Algorithm

367

Table 1. Experimental estimation of the weights of digraph arcs of the tourist network Type of tourism

Holidays in the mountains

SPA tourism

Shopping tours

Ecotourism

Designation

v1

v2

v3

v4

Arc weight, %

42

18

23

17

Fig. 1. Initialization of arcs in the network of the travel services digraph

In view of the global risks of international tourism, tourist service providers should not rely only on one development area or one tourist product. Travel services should be diversified. The sudden occurrence of restrictions associated with outbreaks of coronavirus, natural disasters, armed conflicts, etc. forces to switch to another branch of the network, for which additional vertical ribs are provided in the network between the main routes of the offered services from the sources to drains with their own probabilities of manifestation and the carrying capacity of tourist financial flows. Since the change in the situation is unpredictable, the transition from one branch to another can occur spontaneously, which is reflected in the presence of inclined edges in the original network that have their own direction and their own weight coefficients. Any branch of the components of the offered services can be used as the main branch in the network of the offered tourist product realization. The stopping of the consumption development of this tourist service is carried out synchronously with the moments of the onset of the coronavirus and can also be made at any iteration of the process of implementing a tourist service at any time. Therefore, there is a combination of possible implementations of a tourist product. Due to the fact that the implementation of such a modified product occurs in various places of its use: mountains, SPA tourism, shopping tours, eco-tourism, etc., such combinatorics will be spatial, and in case that the transition from one iteration of the Ford-Fulkerson algorithm to another occurs at different time intervals associated with the onset of COVID-19, such combinatorics will be temporary. The complex use of both types of combinatorics determines its purpose in the implementation of tourist services.

368

M. Sharko et al.

An example of possible implementations of the travel offers combination and their use during coronavirus outbreaks is shown in Figs. 2, 3, 4, 5. In the example No. 1 (Fig. 2), it is impossible to rest in the mountains due to the increasing rains and threats of mudflows on the 5-6th day of the tour, SPA-complexes are closed due to the introduction of mandatory sanitary days due to the coronavirus on the 2nd and 4th days of the tour, shopping trips are not possible due to public holidays on days 3 and 5, and eco-tourism is not possible on days 5–6 due to a storm warning. According to the data received from the Tour PLAZA travel agency, the average tour price is $ 600. Analysis of Fig. 2 shows that if they move along 1 branch, then the travel agency will be able to receive only 4/7 of the planned profit, and 3 days of rest will have to be compensated and returned to tourists. But the ability to switch to another branch allows them to save the entire cash flow.

Fig. 2. Graph for finding the maximum tourist flow in the modified network (example No. 2)

In the example No. 2 (Fig. 3), it is impossible to rest in the mountains due to the increasing rains and threats of mudflows during the first 3 days of the tour; SPA complexes are closed due to the introduction of mandatory sanitary days because of the coronavirus on the 6th day of the tour, shopping trips are not possible due to the public holidays on 5 and 7 days and eco-tourism is not possible due to a storm warning on 2–3 days. If you choose a vacation in the mountains v1 as a priority service, then in order to save profit when selling a tourist product due to changes in weather conditions, you should switch from the main branch v1 to other branches of the offered tour v2, v3, v4, moving from the source s to the drain t along the open paths reflected in the network Fig. 3 taking into account the time constraints on the consumption of this type of service. If we take the services of the offered SPA tourism package (branch v2) as a priority, then the possible path indicated by nodes 1, 4, 6, 9, 13 will be interrupted on the 6th day. Then it will be possible to switch to other types of services v1, v3, v4. If you choose shopping tours as the main priority v3 service, then due to the restrictions caused by outbreaks of coronavirus, for 5–7 days of travel, a possible way for the implementation of this type of service will pass through the nodes 2, 5, 7, 10, after which you can go to the 5th day of branches v1, v3, v4 and on the 6th day again return to node

Combinatorics of the Ford-Fulkerson Algorithm

369

Fig. 3. Graph for finding the maximum tourist flow in the modified network (example No. 1)

16, completing the remaining last unsuccessful for the implementation of services v3, the transition to services v1, v2, v4. If you choose v4 as the main service, the transition to other types of services due to restrictions on the second day of travel is possible on v2 and v3, since the v1 direction is also closed at this time. The service implementation path is indicated by nodes 3, 5, 7, 11, 14, 17, 20. Such a methodology for the implementation of tourist services that are part of a tourist product allows you to keep the profit of manufacturers while meeting consumer demands. Other options for this journey are discussed below. In the example No. 3 (Fig. 4) it is impossible to rest in the mountains due to the increasing rains and the threat of mud flows on the 2nd, 4th, 6th day of the tour, SPAcomplexes are closed due to the introduction of mandatory sanitary days due to coronavirus on the 1st, 3rd, 5th and 7th day of the tour, shopping trips are not possible on days 2, 4, 6 due to public holidays, and eco-tourism is not possible on day 3, 5 due to a storm warning.

Fig. 4. Graph for finding the maximum tourist flow in the modified network (example No. 3)

370

M. Sharko et al.

In the example No. 4 (Fig. 5) it is impossible to rest in the mountains due to the increasing rains and threats of mudflows on the 2nd, 4th, 6th day of the tour; SPAcomplexes are closed due to the introduction of mandatory sanitary days from - due to the coronavirus on the 3rd, 5th and 7th day of the tour, shopping trips are not possible on day 3, 5 due to public holidays, and eco-tourism is not possible on days 5, 6, 7 due to a storm warning.

Fig. 5. Graph for finding the maximum tourist flow in the modified network (example No. 4)

A distinctive feature of the methodology for applying the Ford-Fulkerson algorithms to the task of preserving financial flows when implementing tourist offers in the conditions of COVID-19 is not a decrease in the flow of tourists on the main branches of the network and subsequent iterations on others, but its termination and redirection of the main flow of tourists to other possible offers.

6 Result and Discussion The economic effect of introducing spatial-temporal combinatorics in the implementation of tourist services in context of COVID-19 is to maximize the saving of production costs. In terms of the bandwidth of the edges of the Ford-Fulkerson algorithm, it looks like this: – we choose a travel offer with the maximum attractiveness of characteristics and consumer reviews in the offered package of travel services; – we determine the total value of the bandwidth of the edges of the selected travel service of the network of the digraph of travel offers in normal conditions of the functioning of enterprises; – we establish the path of passage and implementation of this tourist service from source to drain, taking into account the transition to other types of services in connection with the restrictions of COVID-19; – we determine the total value of the bandwidth of the ribs on the chosen path for the implementation of the transformed initial service due to the COVID-19;

Combinatorics of the Ford-Fulkerson Algorithm

371

– we calculate the difference between the sum of the bandwidths of the paths of the initial main service without COVID-19 and the sum of the bandwidths of the transformed services when passing from drain to drain.

7 Conclusion In a stable situation without an emergency travel interruption, the stability of the operation of tourist companies is covered by insurance and stabilization funds, the purpose of which is to compensate for the costs of possible deviations from the offered services, to replace them with other services of similar quality and, in extreme cases, to refund. The risk of such a situation is fully transferred to the entire tourism product. The introduction of the method of switching from one tourist service to another at the time of its termination due to COVID-19 ensures the distribution of risks between the components of the tourist product. The use of combinatorics of the offered services in the space of states, their implementation at different moments of iterations, their application with the synchronization of the throughput of financial flows allows us to reduce the risks in the acquisition of a tourist product and its further implementation. A technology risks of organizing international tourism in the context of the COVID19 pandemic reducing has been developed. It makes it possible to determine the total value of the throughput of all graph arcs. Graph theory using the Ford-Fulkerson algorithm is one of the promising areas of analysis and development management under conditions of uncertainty and ambiguity of the external environment. The above examples of the offered methodology for transforming tourist services due to sudden restrictions caused by COVID-19, at any iteration of the time period for the implementation of the selected services, confirm the effectiveness of the measures to maximize financial flows of manufacturers while maximizing the satisfaction of tourists’ needs.

References 1. OECD: OECD Tourism Trends and Policies 2020, OECD Publishing, Paris (2020). https:// doi.org/10.1787/6b47b985-en 2. Fong, S.J., Dey, N., Chaki, J.: An Introduction to COVID-19. Springer Briefs in Applied Sciences and Technology, pp. 1–22 (2021) 3. Fong, S.J., Dey, N., Chaki, J.: AI-empowered data analytics for coronavirus epidemic monitoring and control. In: Artificial Intelligence for Coronavirus Outbreak. SpringerBriefs in Applied Sciences and Technology. Springer, Singapore (2021). https://doi.org/10.1007/978981-15-5936-5_3 4. Tourism Policy Responses to the coronavirus (COVID-19). https://read.oecd-ilibrary.org/ view/?ref=124_124984-7uf8nm95se&title=Covid-19_Tourism_Policy_Responses 5. Sun, J., Zhang, J.-H., Zhang, H., Wang, C., Duan, X., Chen, M.: Development and validation of a tourism fatigue scale. Tour. Manag. 81, 104121 (2020). https://doi.org/10.1016/j.tourman. 2020.104121

372

M. Sharko et al.

6. Lytvynenko, V., Babichev, S., Wójcik, W., Vynokurova, O., Vyshemyrskaya, S., Radetskaya, S. (eds.): ISDMCI 2019. AISC, vol. 1020. Springer, Cham (2020). https://doi.org/10.1007/ 978-3-030-26474-1 7. Indriastuti, M., Fuad, K.: Impact of covid-19 on digital transformation and sustainability in small and medium enterprises (smes): a conceptual framework (2021). In: Advances in Intelligent Systems and Computing, 1194 AISC, pp. 471–476 (2020) 8. Chatkaewnapanon, Y., Kelly, J.M.: Community arts as an inclusive methodology for sustainable tourism development. J. Place Manag. Dev. 12(3), 365–390 (2019). https://doi.org/10. 1108/JPMD-09-2017-0094 9. Sharko, M.V., Doneva, N.M.: Methodical approaches to transformation of tourist attractiveness of regions into strategic management decisions. Actual Probl. Econ. 8(158), 224–229 (2014). https://eco-science.net/downloads/ 10. Sharko, M.V., Sharko, A.V.: Innovative aspects of management of development of enterprises of regional tourism. Actual Probl. Econo. 7(181), 206–213 (2016). https://eco-science.net/ downloads/ 11. Dinitz, M., Nazari, Y.: Massively parallel approximate distance sketches. In: Leibniz International Proceedings in Informatics, LIPIcs, 153, art. no. 35 (2020). https://doi.org/10.4230/ LIPIcs.OPODIS.2019.35 12. Dinitz, M., Nazari, Y.: Brief announcement: massively parallel approximate distance sketches. In: Leibniz International Proceedings in Informatics, LIPIcs, 146, art. no. 42. Cited 1 time (2019). https://doi.org/10.4230/LIPIcs.DISC.2019.42 13. Wichmann, J., Wißotzki, M., Sandkuhl, K.: Toward a smart town: digital innovation and transformation process in a public sector environment. In: Zimmermann, A., Howlett, R.J., Jain, L.C. (eds.) Human Centred Intelligent Systems. SIST, vol. 189, pp. 89–99. Springer, Singapore (2021). https://doi.org/10.1007/978-981-15-5784-2_8 14. Alford, P., Jones, R.: The lone digital tourism entrepreneur: knowledge acquisition and collaborative transfer. Tour. Manag. 81, 104139 (2020) 15. Babichev, S., Peleshko, D., Vynokurova, O. (eds.): DSMP 2020. CCIS, vol. 1158. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-61656-4 16. Alford, P., Jones, R.: The lone digital tourism entrepreneur: knowledge acquisition and collaborative transfer. Tour. Manag. 81, 104139 (2020). https://doi.org/10.1016/j.tourman.2020. 104139 17. Olijnyk, A.P., Feshanich, L.I., Olijnyk, Y.P.: The epidemiological modeling taking into account the COVID-19 dess emination features. Methods and devices of quality control 2020 – N1(44) pp.138–146 (2020). https://doi.org/10.31471/1993-9981 18. Babichev, S., Lytvynenko, V., Wójcik, W., Vyshemyrskaya, S. (eds.): ISDMCI 2020. AISC, vol. 1246. Springer, Cham (2021). https://doi.org/10.1007/978-3-030-54215-3 19. Dash, P., Rahman, M.M., Zohora, F.T.: An alternate simple approach to obtain the maximum flow in a network flow problem. J. Eng. Appl. Sci. 13(Specialissue10), 8270–8276 (2018). https://doi.org/10.3923/jeasci.2018.8270.8276 20. Laube, U., Nebel, M.E.: Maximum likelihood analysis of the ford–fulkerson method on special graphs. Algorithmica 74(4), 1224–1266 (2015). https://doi.org/10.1007/s00453-0159998-5 21. Qu, Q.-K., Chen, F.-J., Zhou, X.-J.: Road traffic bottleneck analysis for expressway for safety under disaster events using blockchain machine learning. Saf. Sci. 118, 925–932 (2019). https://doi.org/10.1016/j.ssci.2019.06.030 22. Cristiano, P., et al.: Electrical Flows, Laplacian Systems, and Faster Approximation of Maximum Flow in Undirected Graphs. http://math.mit.edu/~kelner/Publications/Docs/maxFlow. pdf 23. Carmen, T.X., et al.: Algorithms: construction and analysis = INTRODUCTION TO ALGORITHMS, p. 1296 (2006). ISBN 0-07-013151-1

Combinatorics of the Ford-Fulkerson Algorithm

373

24. Lawler, E.: Network Flows//Combinatorial Optimization: Networks and Matroids (2011, Dover), pp. 109–177 (2011). ISBN 0486414531 25. Dolgopolov, P., Konstantinov, D., Rybalchenko, L., Muhitovs, R.: Optimization of train routes based on neuro-fuzzy modeling and genetic algorithms. Procedia Comput. Sci. 149, 11–18 (2019) 26. Hashnayne, A.: A proposed linear programming based algorithm to solve arc routing problems. Int. J. Math. Sci. Comput. (IJMSC) 6(2), 61–70 (2020). https://doi.org/10.5815/ijmsc. 2020.02.03 27. Nechirvan, B.I., Hariwan, F.M.S.: Vertex connected domination polynomial of some coalescence of complete and wheel graphs. Int. J. Math. Sci. Comput. (IJMSC) 6(6), 1–8 (2020). https://doi.org/10.5815/IJMSC.2020.06.01 28. Bozhenyuk, A., Gerasimenko, E., Rozenberg, I.: The methods of maximum flow and minimum cost flow finding in fuzzy network. In: Proceedings of the Concept Discovery in Unstructured Data Workshop (CDUD 2012) co-located with the 10th International Conference on Formal Concept Analysis (ICFCA 2012), Katholieke Universiteit Leuven, Leuven, Belgium, pp. 1–12 (2012) 29. Hardesty, L.: First improvement of fundamental algorithm in 10 years, MIT News Office (2010) 30. Gunay, Y. Iskandarli, G.Y.: Applying clustering and topic modeling to automatic analysis of citizens’ comments in e-government. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 12(6), 1–10 (2020). https://doi.org/10.5815/ijitcs.2020.06.01 31. Atwa, W., Almazroi, A.A.: Active selection constraints for semi-supervised clustering algorithms. Int. J. Inf. Technol. Comput. Sci. (IJITCS) 12(6), pp.23–30 (2020). https://doi.org/10. 5815/ijitcs.2020.06.03 32. Sharko, M., Shpak, N., Gonchar, O., Vorobyova, K., Lepokhina, O., Burenko, J.: Methodological basis of causal forecasting of the economic systems development management processes under the uncertainty. In: Babichev, S., Lytvynenko, V., Wójcik, W., Vyshemyrskaya, S. (eds.) Lecture Notes in Computational Intelligence and Decision Making. ISDMCI 2020. Advances in Intelligent Systems and Computing, vol. 1246. Springer, Cham (2020). https://doi.org/10. 1007/978-3-030-54215-3 33. Takahashi, T.: The simplest and smallest network on which the Ford-Fulkerson maximum flow procedure may fail to terminate. J. Inf. Proc. 24(2), 390–394 (2016). https://doi.org/10. 2197/ipsjjip.24.390

Authentication System by Human Brainwaves Using Machine Learning and Artificial Intelligence Z. B. Hu1

, V. Buriachok2

, M. TajDini2

, and V. Sokolov2(B)

1 National Aviation University, Kyiv, Ukraine 2 Borys Grinchenko Kyiv University, Kyiv, Ukraine

{v.buriachok,m.tajdini,v.sokolov}@kubg.edu.ua

Abstract. Authentication and authorization are an indispensable piece of security in computer-based frameworks. As an option for biometrics, electroencephalography (EEG) authentication (authorization) presents focal points contrasted with other biological qualities. Brainwaves are difficult to reproduce, and diverse mental undertakings produce various brainwaves. This examination researches the parts of execution and time-invariance of the EEG-based confirmation. Two arrangements of trials are done to record EEG of various people. We actualize the utilization of artificial intelligence (AI), for example, support vector machine (SVM) and deep neural network (DNN) to characterize EEG of subjects. The correlation between EEG highlights, anodes position, and a mental errand is made. We accomplish more than 90% order exactness utilizing three kinds of highlights from four electrodes. Information from prior meetings is utilized as AI preparing information and information from later meeting are grouped. We discovered that characterization precision diminishes after some time, and inactive undertakings perform in a way that is better than dynamic errands. Keywords: Human brainwave authentication · Biometric authentication · Machine learning authentication · Electroencephalography · EEG · Deep neural network · DNN · Support vector machine · SVM · Keras Neural Network · KNN

1 Introduction To accomplish the examination target of understanding an EEG-based authentication framework, two significant parts of a biometrics validation should be satisfied [1]. They are execution and perpetual quality [2]. From an execution viewpoint, we center around raising the grouping precision of various people. In the perpetual quality angle, the impact of time on the order exactness is examined [3]. We held trials to assemble information and test our proposed EEG highlights and philosophy. In the analyses, we request that subjects perform mental undertakings while EEG is recorded. At that point, we preprocess the crude EEG and concentrate EEG highlights. The EEG highlights are utilized for AI and machine learning (ML) preparation and order. We look at changed mental errands, a mix of EEG highlights, a mix of anodes, and AI/ML conditions [4]. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 374–388, 2021. https://doi.org/10.1007/978-3-030-80472-5_31

Authentication System by Human Brainwaves Using Machine Learning

375

Section 2 provides an analysis of the sources. Section 3 contains the research methodology and indicates the direction of research. Measuring instruments, testing process, and experiment sequence are described in Sect. 4. Results of conducted experiments are given in Sect. 5. The paper ends with acknowledgments and conception of the future work in Sect. 6.

2 Review of the Literature In the last few years have been carried several studies to achieve EEG-based authentication [5]. Despite that, some researchers suggest conditions and factors for the authentication process. Factors include mental task, EEG characteristics, electrode positions, and method of classification. Some studies use ML for the classification method [6]. The study by Ashby et al. [7] uses SVM to classify the EEG characteristics of five subjects. From the EEG recordings, four EEG characteristics are extracted, which are autoregressive coefficients, power spectral density, spectral power, and interhemispheric power difference. Using the SVM classification, the results are encouraging, the classifications produce an average of 97.69% accuracy across four activities. Another study by Hu et al. focuses on EEG-based authentication in a pervasive environment [8]. For a real implementation, it is essential to have a short measuring time and ease of use of the EEG measuring device. The recording device is a cap with a single electrode located in the center of the head [9]. The experiment involved 11 people. Subjects are asked to be inactive during EEG recording. After filtering and noise removal, the characteristics of the coefficients of the autoregressive EEG model are extracted. In this study, the classification of functions does not use ML but uses a naive Bayesian classifier. The obtained results regarding classification demonstrate that the accuracy of the classification depends on the EEG recording timing. The longer is the recording time produces greater accuracy. The minimum accuracy is 66.02% using a sample time of 4 s and the maximum is 100% with a sample time of 56 s. Both studies show classification results for a single recording session, so the effect of time on classification performance is unknown. Another study by Blondet et al. addresses the issue of persistence of EEG-based authentication [10]. The taken experiment lasted five months and consisted of three sessions, which were recorded at different times. There were nine participants in all three sessions. The considered feature of the EEG is event-related potential (ERP) from the mid-occipital electrode. The method of classification used is cross-correlation. The gained results of the experiment demonstrate that significantly high accuracy was maintained between sessions. The average for the second session is about 90% and the average for the third session is about 80%. However, not all outcomes are highly accurate. Only a few subjects achieved an accuracy of less than 70% in some sessions, and one subject only achieved an accuracy of 10% in a given session. This shows that performance is also influenced by individual subject factors [11]. For some people, their brain waves may not change significantly over time, but for some people, the brain waves may change. In [12] author team used a single channel which according to our researches it would not be enough as Tables 7, 8, 9 and 10 of this paper prove more sensors bring more

376

Z. B. Hu et al.

accuracy and taking samples with one channel can cause may fault or false-positive on ML. Unfortunately, in our tests, we couldn’t reach the true-positive rate this paper got but the SVM algorithm has a good result for our tests as well while in paper [13] researchers using Artificial Bee Colony or Practical Swarm Optimization and some more algorithms but in our tests, these algorithms didn’t work as good as DNN and SVM were working but as in [14] shown it’s possible to make a classification based or brain waves and in different cases and situation it might work to group emotions and even include the valence level [4]. Also getting deep into [12] and [15] somehow possible to find this connection that one second samples are not accreted and even signal delay also may influence the results, so we need more what is the most efficient time we still don’t know and it might be part of future works, and this should be the last because firstly the method of taking brain waves and other details should fix and then measure the best time duration for taking signals [16]. Technical solutions are presented in [17] and [18].

3 Research Methodology We held two arrangements of examinations with various number of subjects and mental errands. The primary investigation has the objective of contrasting components locating the best mix for arrangement precision. In this examination, subjects partook in one meeting of recording. Subjects partook in more than one meeting of recording with time stretch. Although they are isolated tests, a few explore conditions are applied to the two investigations. They incorporate the EEG estimating gadget utilized, EEG recording conditions, and trial stream [19, 20].

4 Setting up the Experiment 4.1 Measuring Instruments We use a wearable four anodes EEG estimating gadget called Muse (see Fig. 1), the cerebrum detecting headband made by organization InteraXon [21]. It has the structure of a headband and doesn’t demand conductive gel during recording measures. Muse records EEG and communicates the information to the PC using Bluetooth. This device uses Bluetooth 2.1 with EDR, references Electrode are FPz and it works with 220 or 500 Hz sample rate. The gadget has four chronicle anodes which are two cathodes on the brow and two terminals on the rear of the two ears. The cathodes are named by the 10–10 arrangement of terminal position. They are silver electrodes FP1, FP2 for left and right brow/forehead, and conductive silicon rubber made electrodes TP9, TP10 for the back of the left and right ear (see Fig. 2). During the examinations, crude EEG with a testing rate (500 Hz) is recorded by each of the four terminals and sent to a PC. 4.2 Conditions for EEG Measurement During recording the brainwave of people, although the signs of each continuous mind exercises are recorded too. So to get the brainwave of just the psychological errand

Authentication System by Human Brainwaves Using Machine Learning

377

Fig. 1. Muse band version 2.

Fig. 2. Muse sensor position.

given is exceptionally troublesome. Other mind exercises can be stirred up in the EEG recording. To restrict the proportion of other psyche signals, during the assessment’s subjects are given rules of what not to do during EEG recording meetings (Fig. 3). To decrease the cerebrum signals brought by muscle action, subjects are approached to make their body development as least as could be expected under the circumstances. Yet, imperative body developments, for example, breathing are required by people so they are not denied. Before the chronicle meeting begins, subjects are approached to plunk down in an agreeable position. The psychological undertakings of our examinations don’t need the subjects to watch any picture or recordings so they are approached to close their eyes during recording. This is to decrease the impact of cerebrum signal brought my mind preparing of vision from the feeling of sight. During the chronicles, commotions, and sounds from surround are also taken into account. The experiment room was keeping safe and under control. The individuals inside the room are mentioned to not talk and make uproarious commotions during the

378

Z. B. Hu et al.

EEG recording measure. One of the psychological errands is for subjects to tune in to music from earphones so for this situation sounds from the room don’t have a lot of effects.

Fig. 3. Flowchart of measurement.

In a meeting of trial, a crude EEG information of a subject while performing separate mental assignments is recorded. We begin the meeting by disclosing and offering directions to the subject. The progression of tests and distinctive mental errands are clarified. Then another subject tried to wear the Muse. Since Muse has dry cathodes, the subject can simply wear it with no conductive gels. The situation of Muse is checked so the cathodes are in the right situation on the subject’s head. Hair should be covered to not disturb electrodes. For those who wear glasses, they asked to take their glasses since it will impede the cathodes behind the ears. The deliberate EEG is shown on the PC so, in the event of a lot of commotions, the situation of Muse can be balanced. After Muse was connecter correctly, the subject is approached to remain still for some time. Dream needs an ideal opportunity for the alignment of brainwave estimation. The necessary time is around one minute. The alignment lessens commotions and old data recorded by the gadget. But not all clamors are hindered so during the inclining of crude EEG, separating is applied. In the tests, several mental assignments in the meeting change between first and second tries. The EEG recording will be firstly repeated and then continued. The subject is approached to play out a psychological errand while Muse records the EEG. Recording time for each psychological undertaking is four minutes. In information handling, the first and most recent 30 s are taken out so there is three minutes’ worth of information for each recording. After each psychological errand recording, the subject takes a brief break. During the resting time, the EEG recording is affirmed. The account has no blunders we proceed onward to record the next assignment, for some subject which EEG Data was not enough to clear the recording was repeated. After the whole mental assignment accounts are done and no issues are discovered then the meeting closes [9].

Authentication System by Human Brainwaves Using Machine Learning

379

4.3 Data Testing Process In our assessment, we endeavor to investigate EEG-based authentication by using AI/ML (Fig. 4). Via preparing a model to think carefully, the model perceives novel examples of each subject. In the verification measure, the subject’s introduced certification (as brainwave) is ordered by the model into the various subject classes. To improve characterization precision, there are a lot of elements that can be thought of. We validate various mixes of the mental undertaking, EEG highlights, cathodes, and sort of AI.

Fig. 4. Data testing process diagram.

To test the different factors and get the arrangement exactness, above all else crude EEG information of subjects is gathered from the test. At that point, that crude EEG information is preprocessed to eliminate commotions and antiquities. Then, feature extraction from EEG records will proceed. The obtained EEG highlights information is divided into two datasets. The first part is preparing a dataset that is used to prepare the AI/ML model. The second one is represented by the forecast test dataset which is the contribution for the AI/ML arrangement. The ready model is utilized for order measurement. Predictions also are classified. Based on grouping, we can figure out the level of information which is ordered into the right classes. This is the grouping exactness. 4.4 Experiment Sequence Two tests are performed in this arrangement. For each errand, three minutes’ worth of EEG recording information is gathered. The two undertakings are loosening up errand and listening to the music test [22]: Test 1. Relaxation. Subjects are asked to not move as much as possible to be calm while closing the eyes. They are told to center on their breathing and to not pay attention to any of their thoughts. Additionally, they are asked to minimize their body movements. Room condition is peaceful so external obstructions are limited. This errand is intended to get the brainwaves during a relaxation state.

380

Z. B. Hu et al.

Test 2. Listening to music. Subjects were listening to music with a headphone. The music is instrumental in classic/traditional music where the violin plays the primary instrument [22]. In this undertaking, subjects are additionally approached to close their eyes and produce negligible body movements. We pick the two mental assignments since they are anything but difficult to do and repeat even after time had passed. Since validation is acted in a significant period, clients ought to have the option to be in a comparable state in each verification cycle. These two errands are common and clients don’t have to recollect any guidelines. If we use too complicated undertakings, clients may not remember the inclination when they played out the assignment previously. Each subject was interested in one meeting of examination. There are 25 subjects, which are a compound of 15 males and 10 females. The normal age of all subjects is 25 years old with a standard deviation of 1.90. Information is gathered from four anodes. 4.5 Data Acquisition Method There are three types of EEG characteristics: DFT, ZCR, and Hjorth parameters. By comparing the results of using each type of function, DFT gives the highest classification accuracy, approximately 82% to 89% for both problems and ML. In second place are Hjorth parameters with an accuracy of about 58%, and in third place is ZCR parameters with an accuracy of 35% to 39%. Possible, that the reason for this result is that the quantity of characteristic data in each type is different. DFT has 180 functions, Hjorth has 12 functions and ZCR has four functions. With different amounts of data, ML can learn more from higher dimensional data. In comparison to others, DFT is much higher effective. We outline DFT frequencies from 1 to 45 Hz. In this range, five types of brain waves; delta, theta, alpha, beta, and gamma waves are found at these frequencies [23, 24]. The proportion of the five waves and the narrowness of every wave can be studied by ML models during the experiment’s mental task, which is investigating brain activities. Due to many factors included in the DFT feature, it can be a possible reason which causes very high accuracy during the usage of only DFT. Combining two EEG functions, the combination of DFT and ZCR or Hjorth parameters slightly increases accuracy compared to a single DFT. Using a combination of the ZCR and Hjorth parameters provides an accuracy of approximately 62% to 67%. While it’s an improvement over using both, it didn’t live up to the DFT feature. With a connection between all three EEG features, we can obtain accuracies in a range of 86% to 94%. The incensement from the usage of two features to three is called to be minimal in the SVM classification. Anyway, there is a considerable improvement in accuracy in the DNN classification, which results in 94% for the relaxation task. Comparing the two tasks, in almost all classifications, the relaxation task offers better precision in comparison to the music task. It approves that different tasks have different unique patterns in brain waves. Additionally, the two tasks showed not a big difference—around 7%. Overall, DNN has a better performance in classifying the brainwaves of the subjects compared to the SVM. It is easier to recognize the difference in the classifications, which were taken with DFT. This characteristic includes a large range of qualities, with the

Authentication System by Human Brainwaves Using Machine Learning

381

features of 180 DFT at 45 frequencies from four electrodes. Due to the variety dimension of the function, DNN may be able to extract high-level functions and find more detailed patterns. Studying and confronting various EEG features, we found that the combination of the three proposed types of features gives the best classification accuracy. However, combining DFT with ZCR or Hjorth parameters only gives slightly less precision, so using two EEG functions may be an option if less processing time is required. The recording device Muse EEG has four electrodes: TP10 (right ear), TP9 (left ear), FP1 (left forehead), and FP2 (right forehead) in Fig. 2. We classify subsets of EEG features that are extracted from the brain waves recorded by electrode combinations. In all classifications were used all three types of EEG qualities. According to the obtained results, an escalation in classification accuracy appears when more electrodes were used. This makes sense because people’s brain waves are unique in not just one but many parts of the brain. With more electrodes, you can measure the uniqueness of brain waves from more parts of the brain. However, more electrodes result in more time and difficulty in wearing an EEG recording device. Therefore, it is important to find the best number of electrodes and registration points. During our experiment, we came by results that using four electrodes gives us at least 87% accuracy in DNN. While, using three electrodes gives slightly lower accuracies, around 86% to 91%. The increase in accuracy by using three electrodes on four electrodes is not as significant compared to using two electrodes on three electrodes. Thus, using many electrodes to improve accuracy may not be worth the disadvantages of device cost and wear preparation time. Set on the four individual electrode points, we noticed that the accuracy varies from one to another. The electrodes TP9, FP1, and FP2 bring accuracies of almost 58% but TP10 shows higher accuracies of about 67%. The brain waves of this electrode, therefore, contain clearer patterns compared to the other three points. According to brain anatomy, the TP10 electrode is located very close to the right temporal lobe, which is responsible for dealing with non-verbal memory and reaction like shapes and sounds. Also, it controls social skills and basic behavior. The brain activity of this part may be more exclusive among individuals, resulting in greater classification accuracy. If we use any two electrodes in any combinations, the accuracy still expands between the two chosen points. However, we get similar classification accuracies from using any combination of three electrodes. In this way, there is the possibility to reduce the bias, which was caused due to the selected certain point of the electrode by using more electrodes. When comparing the two ML algorithms, that are using one or two electrodes, the accuracy varies, and neither SVM nor DNN offers better overall performance than the others do. Nevertheless, by using three and four electrodes, DNN produces greater accuracy in all classifications. This could be because DNN can extract higher-level functionality from multiple data dimensions. The result of two mental tasks shows that the relaxation task is better than the music task in almost all classifications. The diversity is not so big if we use a combination of three or four electrodes. Despite that, a big difference can be seen in the ratings using one

382

Z. B. Hu et al.

or two electrodes. The source of this difference is the significant diversity in precision between the relaxation task and the music task of the FP1 and FP2 electrodes. In the classifications, which contain one or the other of the electrodes FP1 and FP2, the performance of relaxation and the mental task is different. In the individual electrode classifications of FP1 and FP2, the relaxation task has higher accuracies than the music task. The diversity is quite large-scale 16–20%. FP1 and FP2 electrodes are based on the forehead, they show that during the subject performing a relaxation task the brain wave near this part is unique. The part of the frontal lobe of the brain is responsible for problem-solving and motor activities. During concentration on listening to music, people don’t think about other stuff, that’s why the frontal lobe brain activity will be decreased. In case when subjects are involved with the relaxation task, they may think about some other things. In this way, the frontal lobe can be more active. Such activity provides more clear EEG features for extraction, for example, the amplitude range of the brain waves. In this series of experiments, we compare several factors in the classification of brain waves. The best precision is obtained in the DNN classification of three types of EEG resources (DFT, ZCR, Hjorth) with four electrodes (TP9, FP1, FP2, TP10) while the subject is performing the relaxation task. By classifying the brain waves of 20 individuals, we achieved a classification accuracy of 94%. The time-invariance viewpoint was likewise tried. It was captured in two different meetings. From the outcomes, we discovered that the exhibition corrupts after some time. We think about three distinctive mental errands. At the point when subjects perform latent errands, more EEG attributes of people can be recognized in later meetings, bringing about better characterization. In executing EEG-based verification, automatic undertakings may be better since subjects can simply react normally to the introduced upgrades. After finding out suitable factors for subject classification, we evaluate the property of time-invariance. Users apply authentication systems for a long period since every time users want to access a computer system, authentication is needed. Noticeably, it is necessary to have a stable and reliable performance over time. To test the property of time invariance, we train and classify ML using two datasets recorded at two different times (Fig. 5). There are two sessions of experiments with an interval between them. The data recorded in Session 1 is used for the training process, and the data recorded in Session 2 is used for the classification process.

5 Experiment Results 5.1 Data Format For each psychological errand, 3 min of crude EEG information is gathered. Before EEG highlights extraction, firstly the preprocessing will be done to delete some noises and old data. The test room has electronic gadgets, for example, PCs forced air systems and lights that produce resistance in the EEG recording. To reduce the resistances, recorded signs of frequencies that are different from the EEG frequencies will be deleted. The strategy for commotion evacuation is used by applying the passband channel to the crude

Authentication System by Human Brainwaves Using Machine Learning

383

Fig. 5. Learning and classification ML diagram.

EEG recording. We used 1–45 Hz FFT (Fast Fourier Transform) channel. Everything lower than 1 Hz and higher than 46 Hz was removed from the results. The remaining signs in the chronicle are the brainwaves which are obtained through EEG highlights extraction measurement. All 180 s recorded brain waves were divided into one-second sections of information and EEG highlights are determined dependent on that window size. It brings about 180 information lines being gathered from every EEG recording. There are three kinds of EEG includes that are extricated from each section: 1. There are three Hjorth Parameters (Hjorth) boundaries: complexity, activity, and mobility. Three Hjorth boundaries highlights are removed from each fragment for a sum of 12 Hjorth features, from four cathodes. 2. Discrete Fourier Transform (DFT) of frequencies 1–45 Hz was recognized from each Sect. 45 DFT highlights are gathered for every anode. Since there are four cathodes, 180 DFT features are accessible for one second EEG fragment. 3. One Zero-Crossing Rate (ZCR) include is processed for each portion. From each of the four cathodes, there are four ZCR features. Each dataset contains 4,500 record lines from 25 subjects. All columns are 196 because of 180 DFT features, 4 of ZCR, and 12 of Hjorth features so in total it is 196. An outline of the entire dataset has appeared previously. Relaxation tests and tests of listening to music have a similar piece of information. Train the model and characterization measurement, we pick subsets of information to analyze various features, cathodes, and tests. 5.2 Usage of Artificial Intelligence and Machine Learning Computer-based intelligence ML is applied to EEG-based authentication in seeing solitary ascribes of each customer’s EEG features. Initially, an AI/ML model is prepared to utilize preparing information on EEG highlights. The model learns the attributes of every client. During verification measures, EEG highlights are separated from EEG tests given by the client, and those highlights are grouped by the AI/ML model. If the level of highlights that are ordered into the right class of subject and reached a specific

384

Z. B. Hu et al.

edge, at that point that subject is validated and obtained access to the framework. In this investigation, we get a dataset of EEG highlights from 20 subjects. The dataset is separated into two sets for AI/ML preparing information and AI/ML grouping forecast test information. Before isolating, we standardize the information utilizing Gaussian standardization. From the 4500 lines, 80% of them are chosen as preparing datasets and 20% as forecast test datasets. The determination is done arbitrarily. Two sorts of AI/ML are utilized in this examination, uphold SVM and DNN. Table 1. SVM data classification sample. Su1 Su2 Su3 Su4 Su5 Su6 Su7 Su8 Su9 Su10 … Su1

28

2

0

0

0

3

0

0

0

0

Su2

0

33

1

0

0

0

1

1

3

1

Su3

0

0

13

2

1

1

0

2

1

0

Su4

0

0

1

12

2

1

0

0

0

1

Su5

0

0

2

1

23

0

0

2

2

2

Su6

0

0

5

3

1

23

1

1

1

0

Su7

0

0

0

0

0

0

26

0

0

0

Su8

0

1

0

1

0

0

0

26

0

0

Su9

2

0

0

2

0

1

0

1

25

0

Su10 0

0

0

0

0

0

0

2

0

30

…

SVM for Calculation Condition. As Table 1 shows, 10-overlay cross approval is performed for each instructional course. This is to guarantee the unwavering quality of the learning cycle. In ordering forecast test information, SVM yields a table that shows the measures of information of each subject class which is arranged into different classes. DNN for characterizations is assembled utilizing DNN execution in Keras Neural Network (KNN) library. Condition for the program is Tensorflow design in Ubuntu OS. While grouping expectation test information, the DNN consequently computes how many percent of information is ordered effectively. This is called the arrangement exactness. By utilizing the two AI/ML calculations, we compute the arrangement correctness of forecast test information. To discover the best blend of EEG highlights, cathodes, and different variables, we select subsets of information measurements and feed them into the AI/ML models.

5.3 EEG Features Accuracies The consequences of the arrangement by the blend of EEG highlights have appeared. Groupings are made utilizing just each component type, the mix of two-element types,

Authentication System by Human Brainwaves Using Machine Learning

385

and the mix of every one of the three-element types. For the orders, we use information from every one of the four cathodes. In every blend, the consequences of utilizing information from loosening up errands and listening to music test are both appeared in Tables 2 and 3. SVM and DNN output have appeared in each table. Table 2. Accuracies per each EEG extracted feature. Test

Hjorth

DFT

ZCR

SVM, %

DNN, %

SVM, %

DNN, %

SVM, %

DNN, %

Relaxation

60.38

58.35

87.86

89.54

39.29

35.40

Music

57.12

56.99

82.56

82.46

36.76

36.01

Table 3. Three EEG extracted features accuracies together. Test

Hjorth, DFT, and ZCR SVM, %

DNN, %

Relaxation 90.43

94.24

Music

91.92

86.94

5.4 Several Electrodes for Classification Accuracies Next, we use subsets of information from various anode sources. There are four cathodes which are TP9 (left ear), FP1 (left brow), FP2 (right temple), and TP10 (right ear). We look at the characterization correctness from various cathodes and blend them. For all groupings, we utilize every one of the three EEG highlights (Hjorth, DFT, and ZCR). Results from relaxation and listening to music test information have appeared in similar output of SVM and DNN appeared in Tables 4, 5, 6 and 7. Table 4. Classification accuracies for each electrode. Test

Electrode FP1

FP2

TP9

TP10

SVM, %

DNN, %

SVM, %

DNN, %

SVM, %

DNN, %

SVM, %

DNN, %

Relaxation

58.35

56.34

53.55

52.86

58.53

56.29

67.92

66.57

Music

40.74

37.96

45.97

38.35

52.35

51.77

66.42

66.38

386

Z. B. Hu et al. Table 5. Two electrodes together classification accuracies. Electrodes SVM, %

DNN, %

SVM, %

DNN, %

SVM, %

DNN, %

SVM, %

DNN, %

FP1 & FP2

DNN, %

TP9 & FP1

SVM, %

Music

TP9 & FP2

DNN, %

xation

TP9 & TP10

81.79

85.49

81.42

86.23

75.97

79.01

81.43

82.61

75.87

83.79

67.88

71.74

80.38

83.13

73.37

78.06

74.92

76.88

75.57

74.46

67.54

71.93

57.04

60.18

Test Rela-

TP10 & FP1

SVM, %

TP10 & FP2

Table 6. Three electrodes together classification accuracies. Test

Electrodes FP1, TP9, & TP10

FP2, TP9, & TP10

FP1, FP2, & TP9 FP1, FP2, & TP10

SVM, %

DNN, %

SVM, % DNN, %

SVM, % DNN, %

SVM, % DNN, %

Relaxation

85.21

92.16

87.33

90.68

84.42

88.58

86.81

91.07

Music

76.71

85.01

83.62

88.11

77.65

82.24

82.03

85.86

Table 7. Four electrodes together classification accuracies. Test

FP1, FP2, TP9, & TP10 electrodes SVM, %

DNN, %

Relaxation 94.20

90.04

Music

86.87

92.70

As those tables are shown in the general DNN algorithm has better performance on our data and seems more accelerated in comparison to SVM abut as results show the SVM algorithm still enough good and trustable. Also in relaxation, we got a better result in comparison to listen to the music but we think this is more because of environment influence while listening to music can have a deep influence also that the eyes are open and visual inputs also added to brainwaves signals and we think if there we can find a solution to use these together while at this stage listen to music has lower accuracy but has better possibilities to work on later. And with 3 min of recording time and two different active tasks, we have better results because giving passive tasks and taking less time for samples to support the exploration of [10] where high identification exactness is accomplished more than a half year utilizing non-volitional EEG.

Authentication System by Human Brainwaves Using Machine Learning

387

6 Conclusions and Future Work In this paper, we investigate two of the indispensable perspectives, execution, and lastingness of the EEG-based validation framework. From the tests, we mentioned a few objective facts about the variables that influence order exactness. High exactness of more than 90% is accomplished by using the proposed three EEG highlights from four-terminal focuses. Singular anode focuses produce changing correctness’s yet by joining EEG caught from different terminals, stable exactness’s can be accomplished. In characterizing subjects, we utilize two kinds of AI, which are upheld SVM and DNN. We discovered that DNN performs better in characterizations with a bigger number of subjects. In applying for genuine practice, ML may be the more appropriate AI. So as previously mentioned in this paper working on extracting more features from listening to music and find a connection to visual inputs at the same time could be one of our future works and make better accuracy. Acknowledgments. This scientific work was partially supported by RAMECS and selfdetermined research funds of CCNU from the colleges’ primary research and operation of MOE (CCNU19TS022).

References 1. Haukipuro, E.-S., et al.: Mobile brainwaves: on the interchangeability of simple authentication tasks with low-cost, single-electrode EEG devices. IEICE Trans. Commun. 102(4), 760–767 (2019). https://doi.org/10.1587/transcom.2018sep0016 2. Huang, H., et al.: An EEG-based identity authentication system with audiovisual paradigm in IoT. Sens. 19(7), 1664 (2019). https://doi.org/10.3390/s19071664 3. Alsunaidi, S.J., Alissa, K.A., Saqib, N.A.: A comparison of human brainwaves-based biometric authentication systems. Intern. J. Biom. 12(4), 411 (2020). https://doi.org/10.1504/ijbm. 2020.10032523 4. Tulceanu, V.: Brainwave authentication using emotional patterns. Intern. J. Adv. Intell. Paradig. 9(1), 1 (2017). https://doi.org/10.1504/ijaip.2017.081177 5. TajDini, M., et al.: Wireless sensors for brain activity—a survey. Electron. 9(12), 1–26 (2020). https://doi.org/10.3390/electronics9122092 6. Chuang, J., Nguyen, H., Wang, C., Johnson, B.: I Think, Therefore I Am: Usability and Security of Authentication Using Brainwaves. In: Adams, A.A., Brenner, M., Smith, M. (eds.) FC 2013. LNCS, vol. 7862, pp. 1–16. Springer, Heidelberg (2013). https://doi.org/10. 1007/978-3-642-41320-9_1 7. Ashby, C., et al.: Low-cost electroencephalogram (EEG) based authentication. IEEE/EMBS Conference on Neural Engineering: 442–445 (2011). https://doi.org/10.1109/ner.2011.591 0581 8. Hu, B., et al.: A pervasive EEG-based biometric system. Int. Workshop Ubiquit. Affect. Awareness Intell. Interac. 17–24 (2011). https://doi.org/10.1145/2030092.2030097 9. Liew, S.-H., et al.: Incrementing FRNN model with simple heuristic update for brainwaves person authentication. IEEE EMBS Conf. Biomed. Eng. Sci. (2016). https://doi.org/10.1109/ iecbes.2016.7843426 10. Blondet, M.V.R., Laszlo, S., Jin, Z.: Assessment of permanence of non-volitional EEG brainwaves as a biometric. IEEE Int. Conf. Identity, Secur. Behav. Anal. 1–6 (2015). https://doi. org/10.1109/isba.2015.7126359

388

Z. B. Hu et al.

11. Johnson, B., Maillart, T., Chuang, J.: My thoughts are not your thoughts. ACM Int. Jt. Conf. Pervasive Ubiquit. Comput.: Adjunct Publ. (2014). https://doi.org/10.1145/2638728.2641710 12. Hagras, S., Mostafa, R.R., Abou Elfetouh, A.: A biometric system based on single-channel EEG recording in one-second. Int. J. Intell. Syst. Appl. 12(5), 28–40 (2020). https://doi.org/ 10.5815/ijisa.2020.05.03 13. Akkar, H.A.R., Jasim, F.B.A.: Intelligent training algorithm for artificial neural network EEG classifications. Int. J. Intell. Syst. Appl. 10(5), 33–41 (2018). https://doi.org/10.5815/ijisa. 2018.05.04 14. Sreeshakthy, M., Preethi, J., Dhilipan, A.: A survey on emotion classification from EEG signal using various techniques and performance analysis. Int. J. Inf. Technol. Comput. Sci. 8(12), 19–26 (2016). https://doi.org/10.5815/ijitcs.2016.12.03 15. Goshvarpour, A., Ebrahimnezhad, H., Goshvarpour, A.: Classification of epileptic EEG signals using time-delay neural networks and probabilistic neural networks. Int. J. Inf. Eng. Electron. Bus. 5(1), 59–67 (2013). https://doi.org/10.5815/ijieeb.2013.01.07 16. Azizi, M.S.A.M, et al.: Authentication with brainwaves: a review on the application of EEG as an authentication method. In: 2018 Fourth International Conference on Advances in Computing, Communication and Automation (2018). https://doi.org/10.1109/icaccaf.2018.877 6850 17. Zhang, X., Yao, L., Huang, C., Tao, G., Yang, Z., Liu, Y.: DeepKey: a multimodal biometric authentication system via deep decoding gaits and brainwaves. ACM Trans. Intell. Syst. Technol. 11(4), 1–24 (2020). https://doi.org/10.1145/3393619 18. Jenkins, J., et al.: Authentication, privacy, security can exploit brainwave by biomarker. Independent Compon. Anal. Compressive Sampling, Wavelets, Neural Net, Biosyst. Nanoeng. XII (2014). https://doi.org/10.1117/12.2051323 19. Seha, S.N.A., Hatzinakos, D.: EEG-based human recognition using steady-state AEPs and subject-unique spatial filters. IEEE Trans. Inf. Forensics Secur. 1–1 (2020). https://doi.org/ 10.1109/tifs.2020.3001729 20. Liang, W., et al.: SIRSE: a secure identity recognition scheme based on electroencephalogram data with multi-factor feature. Comput. Electr. Eng. 65, 310–321 (2018). https://doi.org/10. 1016/j.compeleceng.2017.05.001 21. Gray, S.N.: An overview of the use of neurofeedback biofeedback for the treatment of symptoms of traumatic brain injury in military and civilian populations. Med. Acupunct. 29(4), 215–219 (2017). https://doi.org/10.1089/acu.2017.1220 22. Sjamsudin, F.P., Suganuma, M., Kameyama, W.: A-18–3 experimental results on EEG-based person identification with machine learning (2016) 23. Tangkraingkij, P., Montaphan, A., Nakavisute, I.: An appropriate number of neurons in a hidden layer for personal authentication using delta brainwave signals. In: 2nd International Conference on Control and Robotics Engineering (2017). https://doi.org/10.1109/iccre.2017. 7935076 24. Kumari, P., Vaish, A.: Brainwave based user identification system: a pilot study in robotics environment. Robot. Auton. Syst. 65, 15–23 (2015). https://doi.org/10.1016/j.robot.2014. 11.015

Approximate Training of Object Detection on Large-Scale Datasets Oleksandr Zarichkovyi(B) and Iryna Mukha National Technical University of Ukraine “Igor Sikorsky Kyiv Polytechnic Institute”, Kyiv 03056, Ukraine [email protected]

Abstract. Modern deep learning approaches require an enormous amount of data to produce high-quality robust results, but broad availably of large-scale datasets for computer vision tasks introduced new issues – a long-tail problem and terrifically increased training time. These problems are especially tough for the object detection field where researchers required to use computation heavy neural networks that process high-resolution images to produce state-of-the-art results. Recently researchers concentrated more on solving the long-tail problem as a more promising in terms of detection performance pushing the computational bar even higher. In this paper, we introduced a novel approach for approximated training that significantly reduces training time on large-scale datasets while maintaining competitive state-of-the-art results. The proposed approach has a lot of speedaccuracy trade-offs which allow researchers more precisely calibrate their training pipelines depending on the number of available computational power. We assume that this will allow the research community to validate faster their hypotheses on million-scale datasets and incentivize broader pool of researchers to conduct their research on larger datasets. The proposed approach weights each data sample from the dataset based on an “anomaly” score of deep features produced by the network and selects only a fraction of data samples that have the highest score for the training phase. These samples contain underrepresented classes that contribute more to the model’s convergence compare to a randomly picked sample from the dataset. The proposed approach reduces impact of long-tail problem for unbalanced datasets. We validated the proposed approach by incorporating it into the training pipeline of Cascade R-CNN-based object detector on Open Images V5 dataset which reduced training time from three weeks to one with neglectable degradation of Open Images mAP from 0.60 to 0.59. Keywords: Computer vison · Object detection · Deep neural networks · Supervised learning

1 Instruction Object detection has been a challenging task in the field of computer vision for a long time. It is fundamental and important task for a variety of industrial applications, such as autonomous driving, on-the-go shops, video surveillance, and many others that require a © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 389–400, 2021. https://doi.org/10.1007/978-3-030-80472-5_32

390

O. Zarichkovyi and I. Mukha

deep understanding of a scene. Due to the rapid development of deep learning and broad availability of large-scale datasets, great progress has been achieved in object detection in recent years [1–13]. Open Images Dataset V5 [14, 15] is currently the largest object detection dataset. Different from its predecessors, such as Pascal VOC [16], MS COCO [17], and Objects365 [18], Open Images dataset consists of a million-scale image database. A comprehensive comparison of Open Images with previous existing datasets is shown in Table 1. Open Images includes 12M bounding boxes (b-boxes) for 500 categories which annotated on 1.7M images. This makes Open Images the first public-available million-scale dataset. In the era of deep learning, more training data always benefits the generalizability of model [19]. With the help of a large-scale Open Images detection dataset, the frontier of object detection would be pushed forward a great step. Table 1. Comparison of different publicly available object detection dataset Dataset

Categories

Images

B-boxes

B-boxes per image

Pascal VOC [16]

20

11,540

27,450

2.4

MS COCO [17]

80

123,287

886,287

7.2

Objects 365 [18]

365

638,630

10,101,056

15.8

Open Images[14, 15]

500

1,784,662

12,421,955

7.0

Increased number of images in Open Image dataset requires from 3 to 14 times more computational resources to maintain similar iteration speed compared to predecessor datasets. Previously conducted 3-month long Open Images Challenges 2018 [20] and Open Images Challenges 2019 [21] shown that participants often tend to use undersampling technic and/or utilized a large number of computation resources (up to 512-GPUs clusters) [22–25] to decrease model’s training time and increase their iteration speed which is crucial for short-term competitions. In this paper, we introduced a novel approach for approximated training that significantly reduces training time for large-scale datasets while maintaining state-of-the-art results. The proposed approach weights each data sample from the dataset based on an “anomaly” score of deep features produced by the network and selects only a fraction of data samples that have the highest score for the training phase. These samples contain underrepresented classes that contribute more to the model’s convergence. By incorporating this approach into training pipeline of Cascade R-CNN-based object detector we reduced training time on Open Images from three weeks to one with neglectable degradation of Open Images mAP from 0.60 to 0.59.

2 Related Work In the following section, we will review work on faster training of neural networks and long-tail problem.

Approximate Training of Object Detection on Large-Scale Datasets

391

2.1 Faster Training of Neural Networks In the context of reducing training time, most focus has been put on improving learning rate scheduling [26, 27] or exploiting existing bottlenecks on a hardware side [28]. The most related work to ours is [27] as we concentrate more on the software side of the problem. In [27] authors improved an idea of Cycle learning rate (CLR) policy presented in [26] and developed on it base One Cycle learning rate policy that performs only one cycle of original CLR policy on much higher learning rates which they select via learning rate test. This allows super-convergence of neural networks and reduces training time by order of magnitude which showed by their comprehensive research on MNIST and CIFAR-10/100 datasets. Our work orthogonal to [24] in the way that the introduced approach concentrates more on careful data selection rather than optimization tricks. 2.2 Long-Tail Problem A long-tail problem is a classical issue of any machine learning algorithm on datasets with a large number of categories that often tend to be unbalanced. A machine learning model that optimized by maximum-likelihood algorithms on such datasets tends to show good performance on major categories and underperform minor ones. Recent research work in the Computer Vision area showed that the major reason for performance degradation of minor classes on large datasets with huge categories imbalance, like Open Images [14, 15] (Fig. 1), is actually caused by disbalance between positive and negative gradients [29, 30]. This disbalance rapidly accumulates during

Fig. 1. Long tail-problem on Open Images Dataset [14, 15] – 50 the most frequent categories (or 10% of all categories) cover 84% of all data in the dataset. X-axis represents a zero-based class index of Open Images categories ordered in descending order by appearance frequency. Each blue bar represents the appearance frequency of the category (left Y-axis). Each red dot represents a cumulative sum of images covered by these categories (right Y-axis).

392

O. Zarichkovyi and I. Mukha

training and negative gradients start to overwhelm any positive gradients for such categories which leads to a point when the model simply ignores minor classes to prevent any errors related to them. The most common approach to this problem is to increase the appearance frequency of minor classes – simple oversampling technic. Also, there are more sophisticated technics which based on rebalancing the importance of hard training samples [13, 31] or increasing the importance of positive gradients over negative for underrepresented classes [29, 30]. Previous works concentrated more on reducing degradation of minor categories by artificially increasing their importance in the training process. They rely on object appearance frequency to define major and minor categories. This statistic generally calculated over the dataset before training and fixed for the whole training process. Our work is different from them in the way that the proposed approach relies more on feedback about each sample from the network rather than merely statistical data. This feedback can change from epoch to epoch and hence provide more flexibility for data balancing depending on how well the model performing for each category.

3 Approach In this section, we will first review anomaly detection approach for self-monitoring of deep object detection models proposed in [32]. Then we will introduce the components of proposed approach for faster convergence of deep object detection models. 3.1 Performance Monitoring of Object Detection In [32] was proposed an online anomaly detection module for object detection models whose major purpose was to detect any degradation in object detection performance which may cause serious performance degradation of underlying components of automation systems and hence lead to undesirable behavior, like in autonomous driving software that heavily relies on the good and robust performance of the object detector. Authors proposed to add to R-CNN-based object detectors [4–6] a subnetwork that will use features F = {F1 , F2 , . . . , Fn } generated by the backbone to predict a probability of performance degradation of this model. This subnetwork is trained separately from the R-CNN network by optimizing following objective:  1, mAPI < λ L(I ) = (1) 0, otherwise where L(I ) – loss function for image I , mAPI – performance of object detection model on image I , λ – minimal performance threshold to consider model performance as normal. To improve performance of this subnetwork authors in [32] proposed to fit it not on raw feature maps from the backbone, but on aggregations of these feature maps (which have a spatial size of W × H ) that presented in Eqs. 2–4. H W x=1 y=1 f (x, y) (2) Fmean = W∗ H

Approximate Training of Object Detection on Large-Scale Datasets

Fmax = maxx ∈[1,

H]

maxy ∈[1,

W]

f (x, y)

Fstd = std (f1 ) ⊕ std (f2 ) . . . ⊕ std (fn )

393

(3) (4)

Final input into subnetwork is generated by simple concatenation of these aggregations as shown in Eq. 5. Fmean_max_std = Fmean ⊕ Fmax ⊕ Fstd

(5)

Overview of proposed in [32] architecture is depicted on the Fig. 2.

Fig. 2. The architecture proposed in [32] alert system. The last convolutional feature layer of the backbone is pooled using the mean, max, and std pooling layer to generate an input feature vector for the alert subnet. Alert subnet consists of a binary classifier that predicts the probability of performance drop of the detection network.

3.2 Proposed Approach for Fast Training The approach proposed in [32] can be used not only for detecting performance degradation in online mode, but also for detecting anomalies in feature distributions that cause these performance degradations. By merely re-ranking all images in the dataset by produced performance score we can select for training only those images on which neural network expect the lowest performance. Training on such images will improve the model’s performance greater than images on which model already has good performance. Unlike [32] we proposed to train object detection model and anomaly detection subnetwork simultaneously to prevent storing a massive amount of features on a hard drive to train anomaly detection subnetwork separately. Incorporating anomaly detection network creates neglectable overhead over object detection model and helps to get better feature representation that suitable for both tasks.

394

O. Zarichkovyi and I. Mukha

In order to select the best training samples from the dataset using the proposed approach, we need to estimate an anomaly score for each image in our dataset which can be computationally expensive for large datasets. So, we proposed an approximation of this process. The approximate approach includes an additional stage of pre-selecting top T images for which we will calculate anomaly score and then re-ranking these samples to get top K (K < T ) of them on which we will train our network. This additional stage can be efficiently done by calculating AP on validation Dval dataset and then ranking training images using Eq. 6.    category ∈Categories 1 − APcategory SI = (6) 2i where, SI – ranking score for image I , APcategory – performance of object detection model for category on validation set Dval , i – number times that image I was used for model’s training. Note that we are using the discounting term 2i for image importance which depends on how many times it was already used in the training process. It was done to prevent training on the same images instead of efferently utilize all available data. At the early stages of training backbone weights are changing rapidly, so we cannot reliably solve anomaly detection subtask and hence cannot rely on the prediction of this subnetwork in the data selection process. To solve this issue, we propose to use a warming-up phase on which the object detection network is pre-trained on a full dataset for several epochs to stabilize weights of the backbone. During this warm-up phase of training, we are using simple supervised training on the full dataset. Note, that the proposed approach requires at least one warm-up epoch for anomaly subnetwork to learn meaningful outputs when initialized with random weights. Also, it is crucial for any deep object detection model to properly estimate the real distribution of each class in the dataset, but the proposed approach artificially shifts these distributions, and hence performance of such model can be degraded for major classes. To solve this issue, we add a fine-tuning phase which includes training on a full dataset for several epochs at the end of training to inject information about the real appearance frequency of each category to the network. On high-level proposed approach can be formulated as Algorithm 1. Algorithm 1. Fast-converging supervised learning of deep object detectors Inputs: Set of images I = { I1 , I2 , I3 , . . . , In }, set of corresponding bounding-box labels for these images L = { l1 , l2 , l, . . . , ln } which together form dataset D = {I , L}. This dataset spited into training Dtrain , validation Dval and test Dtest subsets. Number of training epochs is Etrain . Length of pre-training and fine-tuning phase is Epre−train and Efinetune respectively. Number of pre-selected image candidates is T for each epoch of fast training. Number of training images is K for each epoch of fast training. Minimal pre-image detection performance AP threshold is λ (continued)

Approximate Training of Object Detection on Large-Scale Datasets

395

(continued) Algorithm 1. Fast-converging supervised learning of deep object detectors 1. Pre-train object detection model for Epre−train epochs on full dataset // Fast training loops 2. For each epoch in Etrain : 2.1. Get AP for each category on dataset Dval 2.2. Rank each image in Dtrain using Eq. 6 2.3. Select top T images with the highest score from the 2.2 2.4. Calculate anomaly score for each image from 2.3 2.5. Re-rank images from 2.3 using anomaly score calculated in 2.4 2.6. Select top K images with highest anomaly score from the 2.5 2.7. Train detection model and anomaly detection subnet on images from 2.6 3. Fine-tune object detection model for Efinetune epochs on full dataset

4 Experiments 4.1 Dataset For training purposes, we used a subset of Open Images Dataset V5, which was used for conducting Kaggle Open Images 2019 [33] competitions. It includes a subset of 500 most common categories annotated on 1.7M images. For validation purposes, we used a validation subset of 100k images which was recommended by Kaggle Open Images 2019 organizers for local validation. For testing purposes, we used Open Images private test set that was used during Kaggle Open Images 2019 competitions. All tests were done on Kaggle evaluation servers via late submissions to Kaggle Open Images 2019 – Object detection competition [33]. 4.2 Implementation Details All experiments were conducted using MMDetection framework [34]. All experiments were conducted with Cascade R-CNN [9] architecture with FPN [5] and DCN [7]. As feature extractor we used ResNeXt-101-32x8d [35] pre-trained on Instagram hashtags [36]. All models trained with default MMDetection [34] configs for Cascade R-CNN. We used stochastic gradient descent with the momentum set to 0.9 for optimization. The base learning rate is set to 0.04, which decayed to 0.004 and 0.0004 at the end of 9th and 11th epoch respectively. The total number of epochs is set to 12. The batch size is set to 32. During the training phase, we used only image flip augmentation. For proposed approach we used following parameters: Epre−train = 2, Etrain = 9, Efinetune = 1, T = 400000, K = 200000, λ = 0.3 if not specified anything else. 4.3 Results Results of concluded experiments are presented in the Table 2 and Table 3.

396

O. Zarichkovyi and I. Mukha

Table 2. Comparison of proposed approach with other supervised training technics on Open Images V5 dataset Method

Test set Open Images mAP

Training time (GPU-hours)

Training time (wall clock)

MMDetection baseline 0.600 [34]

4024 h

20.9 days

MMDetection baseline 0.540 [34], less iterations

1360 h

7.0 days

Super-convergence [27]

0.558

1360 h

7.0 days

Oversampling [23]

0.578

1360 h

7.0 days

Proposed approach

0.590

1392 h

7.3 days

Table 3. Comparison of proposed approach with other supervised training technics on different subsets of minor, medium and major categories. Column “Index 1–100” lists test score for the 1st to 100th categories ascending ordered by their appearance frequency in the dataset Method

Index 1–100 (minor)

Index 101–350 (medium)

Index 351–500 (major)

MMDetection baseline [34]

0.471

0.687

0.539

MMDetection baseline [34], less iterations

0.381

0.643

0.472

Super-convergence [27] 0.379

0.662

0.505

Oversampling [23]

0.428

0.690

0.489

Proposed approach

0.456

0.688

0.516

As we can see from Table 2, the proposed approach has slightly lower performance compare to MMDetection baseline, but outperform it in terms of training speed by a large margin. In order to verify our findings, we conducted a set of additional experiments. First of all, we reduced the number of training iterations of the MMDetection baseline to verify that our approach yields higher results if a number of iterations is the same. As we can see from Table 2, the model trained with fewer iterations is underfitted and underperform both the MMDetection baseline and proposed approach by a large margin. Secondly, we compared the proposed approach with the super-convergence approach proposed in [27]. We performed an LR test and fixed the maximal learning rate at 0.04 which changed by the cosine learning rate policy over training process. The number iteration is fixed to the same number as for our approach. As we can see the superconvergence approach [27] is underperform ours by a large margin. This related to a low performance for minor categories, as we can see from Table 3.

Approximate Training of Object Detection on Large-Scale Datasets

397

Lastly, we compared our approach with an oversampling approach proposed in [23] as consider our approach as a variation of oversampling technic. As we can see from Table 2, the results of the proposed approach are higher compared to the oversampling approach. As we can see from Table 3, approach proposed in [23] tends to overfit on minor classes and underfit on major ones which leads to a lower performance compared to our approach.

5 Ablation Study In the ablation study section, we will cover three major components of our algorithm – discount term, pre-training phase and fine-tuning phase. 5.1 Discount Term During the visualization of selected images by the proposed approach for training phase we noticed that some images tend to appear more frequently than others. After deeper revision, we discovered that these images often contain minor categories that have lower performance compare to others. Such undesirable behavior may lead to overfitting on images with minor categories. To reduce the impact of such images on the training process, we introduced the discounting term 2i for image importance in Eq. 6. Such term preventing us from training on the same images and allow us efferently utilize all available data. We consider Eq. 6 as a ranking function and utilized 2i term due to its wide usage in ranking algorithms. A comparison of different discount terms showed in Table 4. As we can see from Table 4, exclusion of discount term leads to a huge drop in performance due to overfitting on images with minor categories. Difference between linear and more sophisticated term is minimal. We recommend using 2i due to the efficiency of performing 21i operation on modern computers. Table 4. Comparison of different discounting terms Discounting term Test set Open Images mAP None

0.581

Linear (i)

0.589

  Exponential 2i 0.590

5.2 Pre-training Phase As was already mentioned in Sect. 3, at the early stages of training backbone weights are changing rapidly, so we cannot reliably solve anomaly detection subtask and hence cannot rely on the prediction of this subnetwork in the data selection process. To solve

398

O. Zarichkovyi and I. Mukha

this issue, we proposed to use a warming-up phase on which the object detection network is pre-trained on a full dataset for several epochs to stabilize the weights of the backbone. The dependence of final score from the number of pre-training epochs showed in Table 5. As we can see, a number of pre-training epoch is a simple trade-off between final performance and training speed. We consider two pre-training epochs as an optimal choice in terms of a speed-accuracy trade-off. Also, we remind that our approach requires at least one pre-training epoch for anomaly subnetwork to learn meaningful outputs when initialized with random weights. Table 5. Dependence of final score from a number of pre-training epochs Number of pre-training epochs

Test set Open Images mAP

Training time (GPU-hours)

1

0.584

1112 h

2

0.590

1392 h

3

0.593

1696 h

5.3 Fine-Tune Phase As was already mentioned in Sect. 3, it is crucial for any deep object detection model to properly estimate the real distribution of each class in the dataset. Methods based on shifts of real distributions in favor of minor categories tend to underperform on major categories. To solve this issue, we proposed to add a fine-tuning phase which includes training on a full dataset for several epochs at the end of the training process to inject information about the real appearance frequency of each category to the network. The dependence of final score from a number of fine-tuning epochs is shown in Table 6. As we can see, the number of fine-tuning epochs is a simple trade-off between final performance and training speed. We consider one fine-tuning epoch as an optimal choice in terms of a speed-accuracy trade-off. Table 6. Dependence of final score from a number of fine-tuning epochs Number of fine-tune epochs

Test set Open Images mAP

Training time (GPU-hours)

0

0.586

1112 h

1

0.590

1392 h

2

0.594

1696 h

6 Conclusion In this paper, we presented a novel approach for fast convergence of deep object detectors which weights each data sample from the dataset based on their “anomaly” score of

Approximate Training of Object Detection on Large-Scale Datasets

399

deep features produced by the network and selects only a fraction of data samples that have the highest score for the training phase. These samples contain underrepresented classes that contribute more to the model’s convergence compare to a randomly picked sample from the dataset. The proposed approach reduces impact of long-tail problem for unbalanced datasets. We conducted a set of ablation studies which show that the proposed approach has a lot of speed-accuracy trade-offs that allow more precise calibration of training pipelines depending on the number of available computational power. By using proposed approach, we reduced training time of Cascade R-CNN on Open Images dataset from three weeks to one week with neglectable degradation of Open Images mAP from 0.60 to 0.59. The use of the research results will allow researchers faster validate their hypotheses, speed-up transfer from thousands- to million-scale datasets in object detection field and allow a wider group of researchers without a large number of computational resources to do their research on larger datasets.

References 1. Girshick, R., Donahue, J., Darrell, T., Malik, J.: Rich feature hierarchies for accurate object detection and semantic segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 580–587 (2014) 2. He, K., Zhang, X., Ren, S., Sun, J.: Spatial pyramid pooling in deep convolutional networks for visual recognition. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8691, pp. 346–361. Springer, Cham (2014). https://doi.org/10.1007/978-3-31910578-9_23 3. Girshick, R.: Fast r-cnn. In Proceedings of the IEEE International Conference on Computer Vision, pp. 1440–1448 (2015) 4. Ren, S., He, K., Girshick, R., Sun, J.: Faster r-cnn: towards real-time object detection with region proposal networks. Adv. Neural Inf. Process. Syst. 91–99 (2015) 5. Lin, T.Y., Dollár, P., Girshick, R., He, K., Hariharan, B., Belongie, S.: Feature pyramid networks for object detection. In: CVPR, 1, 3 (2017) 6. He, K., Gkioxari, G., Dollar, P., Girshick, R.: Mask r-cnn. In: Computer Vision (ICCV), 2017 IEEE International Conference on, pp. 2980–2988. IEEE (2017) 7. Dai, J., et al.: Deformable convolutional networks. In: Proceedings of the IEEE International Conference on Computer Vision, pp. 764–773 (2017) 8. Liu, S., Qi, L., Qin, H., Shi, J., Jia, J.: Path aggregation network for instance segmentation. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 8759– 8768 (2018) 9. Cai, Z., Vasconcelos, N.: Cascade r-cnn: Delving into high quality object detection. arXiv preprint arXiv:1712.00726 (2017) 10. Lin, T.-Y., Goyal, P., Girshick, R., He, K., Dollar, P.: Focal loss for dense object detection. IEEE Trans. Pattern Anal. Mach. Intell. (2018) 11. Xuan, N.H., Hu, Y., Amin, M.A., Khan, G.H., Thinh, L.V., Truong, D.-T.: Detecting video inter-frame forgeries based on convolutional neural network model. Int. J. Image, Graph. Sig. Process. (IJIGSP), 12(3), 1–12 (2020) 12. Aung, M.M., Khaing, P.P., San, M.: Study for license plate detection. Int. J. Image, Graph. Sig. Process. (IJIGSP), 11(12), 39–46 (2019).https://doi.org/10.5815/ijitcs.2020.05.05 13. Dalal, A.-A., Shao, Y., Alalimi, A., Abdu, A.: Mask R-CNN for geospatial object detection. Int. J. Inf. Technol. Comput. Sci. (IJITCS), 12(5), 63–72 (2020). https://doi.org/10.5815/iji tcs.2020.05.05

400

O. Zarichkovyi and I. Mukha

14. Benenson, R., Popov, S., Ferrari, V.: Large-scale interactive object segmentation with human annotators. In: CVPR (2019) 15. Kuznetsova, A., et al.: The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. arXiv:1811.00982 (2018) 16. Everingham, M., Van Gool, L., Williams, C.K., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. Int. J. Comput. Vis. 88(2), 303–338 (2010) 17. Lin, T.-Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft coco: common objects in context. In: Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T. (eds.) ECCV 2014. LNCS, vol. 8693, pp. 740–755. Springer, Cham (2014). https://doi.org/ 10.1007/978-3-319-10602-1_48 18. Shao, S., et al.: Objects365: a large-scale, high-quality dataset for object detection. Conf. Comput. Vis. Pattern Recogn. (CVPR) (2019) 19. Mahajan, D., et al.: Exploring the limits of weakly supervised pretraining. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) ECCV 2018. LNCS, vol. 11206, pp. 185–201. Springer, Cham (2018). https://doi.org/10.1007/978-3-030-01216-8_12 20. https://storage.googleapis.com/openimages/web/challenge.html 21. https://storage.googleapis.com/openimages/web/challenge2019.html 22. Akiba, T., Kerola, T., Niitani, Y., Ogawa, T., Sano, S., Suzuki, S.: PFDet: 2nd Place Solution to Open Images Challenge 2018 Object Detection Track. arXiv:1809.00778 (2018) 23. Gao, Y., et al.: Approach for Large-Scale Hierarchical Object Detection. arXiv:1810.06208 (2018) 24. Bu, X., Peng, J., Wang, C., Yu, C., Cao, G.: arXiv:1910.12044 (2019) 25. Niitani, Y., et al.: Team PFDet’s Methods for Open Images Challenge 2019. arXiv:1910. 11534 (2019) 26. Smith, L.N.: Cyclical Learning Rates for Training Neural Networks. arXiv:1506.01186 (2015) 27. Smith, L.N., Topin, N.: Super-Convergence: Very Fast Training of Neural Networks Using Large Learning Rates. arXiv:1708.07120 (2017) 28. Choi, D., Passos, A., Shallue, C.J., Dahl, G.E.: Faster Neural Network Training with Data Echoing. arXiv:1907.05550 (2019) 29. Ben-Baruch, E., et al.: Asymmetric Loss For Multi-Label Classification. arXiv:2009.14119 (2020) 30. Tan, J., et al.: Equalization Loss for Long-Tailed Object Recognition. arXiv:2003.05176 (2020) 31. Shrivastava, A., Gupta, A., Girshick, R.: Training Region-based Object Detectors with Online Hard Example Mining. arXiv:1604.03540 (2016) 32. Rahman, Q.M., Sünderhauf, N., Dayoub, F.: Performance Monitoring of Object Detection During Deployment. arXiv:2009.08650 (2020) 33. https://www.kaggle.com/c/open-images-2019-object-detection/ 34. Chen, K., et al.: MMDetection: Open MMLab Detection Toolbox and Benchmark. arXiv: 1906.07155 (2019). Source code. https://github.com/open-mmlab/mmdetection 35. Xie, S., Girshick, R., Dollár, P., Tu, Z., He, K.: Aggregated Residual Transformations for Deep Neural Networks. arXiv:1611.05431 (2016) 36. D., Mahajan, et al.: Exploring the Limits of Weakly Supervised Pretraining. arXiv:1805. 00932 (2018)

Framework for Developing a System for Monitoring Human Health in the Combined Action of Occupational Hazards Using Artificial Intelligence and IoT Technologies Oleksandra Yeremenko1 , Iryna Perova1(B) , Olena Litovchenko2 and Nelia Miroshnychenko1

,

1 Kharkiv National University of Radio Electronics, 14 Nauky Avenue, Kharkiv, Ukraine

[email protected] 2 Kharkiv National Medical University, 4 Nauky Avenue, Kharkiv, Ukraine

Abstract. The paper is aimed at solving an urgent and important problem of developing a system for monitoring human health in the combined action of occupational hazards using artificial intelligence methods and infocommunications to ensure its resilience and security. It envisages conducting experimental research on animals under the combined occupational hazards, the results of which will be analyzed using modern methods of artificial intelligence, like machine learning and deep learning systems. The architecture of the human health monitoring system will be proposed based on the latest architectural approaches built upon IoT technologies and integrated cloud computing, provided that the resilience and security of such distributed systems are ensured. The result, which will solve this problem, will be the concept of building a monitoring system that differs from existing ones by using artificial intelligence methods and infocommunications, which will allow monitoring of human health in real-time and respond to critical changes. Keywords: Occupational hazards · Infocommunications · Data mining · Machine learning · System analysis · System evaluation

1 Introduction The significant material accumulated by the authors reflects that today, due to intensive and constant changes in the environment, as well as the conditions of their existence, man is exposed to exogenous, biotic, abiotic, natural and anthropogenic adverse environmental hazards, in particular: non-ionizing and ionizing radiation, temperature changes, malnutrition, hyperoxia or hypoxia, pollutants, and other stressors [1, 2]. This number and variety of hazards also implies their possible relationship with each other. In this regard, it is necessary to study the combined effect of physical hazards, the effect of low frequency electromagnetic radiation under cold stress on the immune system, the functioning of the reproductive system, shifts in biochemical processes and changes in the endocrine system under the combined influence of these hazards. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 401–410, 2021. https://doi.org/10.1007/978-3-030-80472-5_33

402

O. Yeremenko et al.

Comparative characterization of the obtained data should be carried out using modern methods of data mining, which will determine which of the studied hazards has the greatest negative effect in each of the studied systems of the body. Considering this, the scientific and applied problem of developing a system for monitoring human health in the combined action of occupational hazards using artificial intelligence methods and IoT technologies and integrated cloud computing under the provision of resilience and security of such distributed systems is relevant. A problem to be solved in the work is the application of modern methods of artificial intelligence to determine the criterion-relevant indicators of the functional state of the body to improve the monitoring of human health, exposed to a set of production factors, based on intelligent data processing using hybrid neuro-phases machine learning systems and determination of leading indicators under the combined influence of factors of the production environment. The object of the work is the process of monitoring human health with the combined action of harmful production factors using artificial intelligence methods and infocommunications in the development of a monitoring system ensuring its reliability and security [3, 4].

2 Related Research Analysis There are a number of scientific developments devoted to the study of changes in the response of a living organism to toxic substances under conditions of exposure to low temperatures and certain pathognomonic conditions under conditions of this effect [5]. The combined action of electromagnetic fields (EMF) of industrial frequency, noise, elevated air temperature in the working area of air traffic controllers, telephone operators of modern digital communication, combined action of production hazards of garment production was studied. When assessing the impact of a set of hazards, including sound, light, and EMF from the monitor, the probabilistic relationship between the action of sound and light near the monitor, as well as the relationship between EMF and light [6, 7]. A number of works are devoted to the study of synergetic biological effects of the combined action of EMR and ionizing radiation [8]. Hygienic, medical and biological studies conducted in various fields of economic activity indicate a set of hazards of the production environment that affect the body of workers. Including the most common complex effects of chemical accompanied by physical hazards, namely noise, air temperature, vibration, radiation of different ranges, etc. [9]. That is, work processes lead to more adverse hazards that become increasingly difficult to identify and assess, the interaction of hazards with each other can cause multidirectional reactions of different functional systems of the body [10]. That is why it is necessary to use intelligent methods to determine the informative value of medical indicators and understand the effects of adverse factors and their combination on the basis of artificial intelligence methods, like machine learning and deep learning approaches [11] and also to use a set of sensors to measure the required indicators, taking into account the conditions in which a person is [12]. At the same time, the purpose of the work [13] was the creation and development of a specialized infrastructure for a smart health care system based on Internet of Things

Framework for Developing a System for Monitoring Human Health

403

technologies (IoT). It was analyzed relevant existing technological standards, communication protocols, software and hardware taking into account the requirements of the IoT system used to design this system. Whereas [14] presents the IoT eHealth architecture that is fault-tolerant to external destructive influences.

3 Basic Approaches and Foundations of System for Monitoring Human Health in the Combined Action of Occupational Hazards The principal idea of the theoretical foundations for developing a system for monitoring human health in the combined action of occupational hazards consists of few parts. 1. The first step is experimental research on animals (laboratory white rats WAG line) that will simulate the real conditions of the production environment. Then the functional state of the animal’s organism under the combined influence of occupational hazards will be studied. The state of animals can be described by a set of medical indicators. 2. In the second step, the modern methods of intelligent data processing based on hybrid neuro-fuzzy systems learn obtained indicators that allow to determine the leading factors influencing together and to establish the criterial significant indicators that will fully reflect the functional state of the organism in these conditions. 3. In the last step, it is planned to develop a system for dynamic monitoring of occupational hazard values and at the same time for simultaneous monitoring of certain informative medical indicators. 4. In addition, it is proposed to use the system both for dynamic monitoring of levels of production factors and for simultaneous monitoring of certain informative medical indicators based on the use of IoT and integrated cloud computing technologies. This approach will allow at future create a system for online monitoring of the occupational hazards and health of workers in order to respond quickly and timely implementation of preventive actions. Therefore, the aim of the work is to advance the theory and practice of human health monitoring in the presence of various occupational hazards. The peculiarity of the work is the use of modern intelligent methods of processing, transmission, and protection of information. The practical part of the work is the future development of the technical tools for staging and conducting experiments on animals, as well as the creation of a monitoring system using artificial intelligence methods and infocommunication means. To achieve this goal it is necessary to perform the following tasks: 1. Conducting the analysis of the impact of existing exogenous, biotic, abiotic, natural, and anthropogenic adverse environmental factors on humans at work. In the beginning, we should identify the most common factors that affect employees in the complex during the performance of professional duties and establish at what levels and in what concentrations occupational hazards are present in the workplace and in the work area. Then determine the timing of the impact of these factors on employees. Based on the obtained results, an experimental model will be developed that will fully reflect the real conditions of the worker’s environment.

404

O. Yeremenko et al.

2. Carrying out of an experiment on animals in the conditions of the combined action of various occupational hazards, the description, and reception of data. This stage consists of investigating the biochemical parameters of laboratory animal’s blood, which will allow observing the nature of functional changes during the experiment. At this stage, it is important to be sure that experimental laboratory conditions and workplace conditions are the same. 3. Performing the study of modern methods of artificial intelligence that can be used for intellectual analysis of experimental data, including machine learning and deep learning methods. 4. Carrying out of the intellectual analysis of the received experimental data with the application of methods of artificial intelligence, the corresponding conclusions, and development of recommendations. At this stage, there is a need to determine the most informative parameters of the organism that have changed during experimental studies and to derivate computational algorithms and software implementation of methods for visualization of the obtained results. 5. Analysis of modern technological solutions using IoT/IoE and cloud computing technological solutions for building a distributed monitoring system. 6. Development of the architecture of the system of monitoring the state of human health with the combined action of occupational hazards using the methods of artificial intelligence and infocommunication means to ensure its resilience and security. Taking all into account, as an approach to solving the problem of developing a system of monitoring human health during production activities with the combined action of several occupational hazards (electromagnetic radiation, temperature, chemical effects, etc.) are used information and communication technologies (IoT and extended cloud computing) for registration, transmission, processing, and protection of biomedical information. The novelty of this approach is a combination of experimental research on animals, which simulates the real conditions of production, intelligent processing of information to determine the most informative changes in functional status, as well as information and communication technologies used to measure, transmit, store and processing of data on the state of human health and occupational hazards. Furthermore, during the intellectual analysis of experimental data, modern methods of machine learning will be used, namely hybrid neuro-phase systems whose advantage is the ability to process information in real-time, when it is presented as a data stream that is sequentially received for processing [15–18]. When determining the informativeness of medical symptoms, the use of a hybrid system Feature Selection-Extraction is proposed. At the same time, the human health monitoring system is based on the use of the latest architectural approaches, built using IoT technologies and extended cloud computing, including Cloud, Fog, and Edge Computing, provided that the integrated architectural and operational resilience and security of such distributed systems [13, 14, 19–25]. Note that the main technological features of human health monitoring system include the following: 1. Identification of the most informative medical characteristics based on the conducted experiment on animals at the combined action of occupational hazards on them.

Framework for Developing a System for Monitoring Human Health

405

2. Design of a human health monitoring system with measurement of a limited number of the most informative medical signs and production factors. 3. Obtaining digital data on the human condition (electronic medical records), their coordination, and exchange between the relevant parties within the monitoring system, which will improve the results of relevant decisions. 4. Mobile access to information that ensures its receipt in real-time, and thus increases the productivity of the monitoring system as a whole. 5. Remote monitoring technologies will promote the use of innovative models, such as remote sensing and diagnostics, which will allow the intellectual analysis of nonstationary time series at an early stage to detect changes in human well-being in the workplace. 6. Generation of hypotheses based on the obtained data using artificial intelligence (AI) methods and Machine Learning (ML) algorithms, including Deep Learning (DL) models. Such an approach combined with analytics provides new insights from data that can support, in particular, evidence-based clinical decisions, personalized medicine, public health management, and clinical trials. 7. The use of cloud technology provides a secure, fast, and cost-effective way to access, store, and share human health information between different stakeholders and multiple connected devices. The flexibility, scalability, and interactivity of cloud computing, as the main hub for information and data exchange, provide greater opportunities to improve diagnostics and monitoring results.

4 Intellectual Analysis of Experimental Data A significant advantage over existing studies is that they are based on the measurement of the most informative medical characteristics, the number of which was determined by intelligent analysis of the parameters obtained during the experiment on animals exposed to the combined effects of various occupational hazards. The expected results are significantly different from previous methods, as all previous experiments took into account either the action of only one harmful production factor or the combined action of two [1, 4]. The presented approach proposes to investigate the combined action of electromagnetic radiation, temperature, and chemical effects simultaneously. Therefore, obtained from experimental data on animals, informative medical indicators that fully describe the change in functional status will determine the specific features that need to be measured in humans, limiting the need to use a large number of sensors. 4.1 Establishment of Hazards and Indicator Sets Suppose that hazards set H is a set of occupational hazards affecting the workers expressed as: H = {h1 , h2 , h3 , . . . , hz }. While the medical indicators set MI is a set of medical indicators measured from laboratory animals and expressed as: MI = {mi1 , mi2 , mi3 , . . . , min }.

406

O. Yeremenko et al.

Among them, some medical indicators are more significant than others and it is necessary to determine it for reflecting the functional state of the organism and for further measurements on humans. 4.2 Feature Selection-Extraction Method Feature Selection-Extraction method [15–18] permits to obtain a reduced number of medical indicators F that are mostly reflecting the functional state of the organism and expressed as: F = {f1 , f2 , f3 , . . . , fm }, where m < n. The purpose of the Feature Selection-Extraction method is the transformation of MI to F according to the information value of medical indicators. Number m can be defined by physician and can change from one group to another. Therefore, it is important to range medical indicators from most informative to fewer ones using a hybrid system for feature selection-extraction tasks, presented in Fig. 1. It contains a module for centering and normalization of medical indicator’s values, covariance matrix detection module, 1-st PC (principal component) calculation module, distance calculation module, and feature reduction module. It should be noted that the module for centering and normalization performs data centering and data normalization using formula:   mii − mii − min(mii ) . ci = max(mii ) − min(mii )

Fig. 1. Structure of hybrid system for feature selection-extraction tasks.

Framework for Developing a System for Monitoring Human Health

407

The covariance matrix detection module calculates the covariance matrix of normalized features ci and returns matrix L that consists of the eigenvectors of the covariance matrix. The module of 1-st PC definition calculates the first principal component like a vector: Pr = l1 c , where l1 is the first row of matrix L, c = (c1 , c2 , c3 , . . . , cn )T is a vector of centered and normalized medical indicators. At the module of distance calculation distances between c and Pr are calculated using the formula:  |Pr − c|. d= n

The smallest distance corresponds to the medical indicator with the highest information value. Then this particular indicator is removed from the dataset at the module of feature reduction and the whole process of calculations is repeated again for a reduced number of medical indicators. It helps to obtain ranged by information value set of medical indicators F = {f1 , f2 , f3 , . . . , fm , . . . , fn }.

5 Technological Features of the Proposed Human Health Monitoring System The developed monitoring system should provide a number of important functions for the stable and efficient operation of the connected relevant medical services: 1. Security, along with confidentiality and data protection, is an extremely important aspect of health care and is the basis of trust in the provision of clinical services. In addition, the latest network technologies support the functionality of information protection, tracking of its transmission, detection of vulnerabilities, and prevention of attacks with the ability to segment traffic when it is transmitted over the network. 2. Automation of monitoring system management provides adaptation and prompt response to changes in the network, as well as comprehensive control and compliance with the latest security and resilience policies. 3. Availability of different classes of services when the priority of traffic associated with a particular health care application, provided that it is ensured that non-critical services do not dominate in the bandwidth consumption to the detriment of services related to monitoring. 4. Traffic monitoring and analysis in order to classify, identify anomalies, and effectively manage network resources.

408

O. Yeremenko et al.

5. Further combination of hybrid neuro-fuzzy systems for intelligent analysis of the registered time series of changes in human medical indicators simultaneously with the indicators of the production environment will allow monitoring of health in real-time. Moreover, the scientific novelty of traffic management methods providing resilience and security, on which the functioning of the system of monitoring human health at work is based, is that, in contrast to existing systems, is a complex technological solution taking into account the properties of cyber resilience and the corresponding architecture.

6 Conclusion The proposed framework is based on the development of a clear algorithm (based on informative evidence) for the study and analysis of the functional state of workers, which will save lives, health, and efficiency in the process of their working. The introduction of modern information and communication technologies will improve the conduct of clinical and hygienic monitoring through a systematic analysis of the working environment and health of workers. It will modernize the principle of occupational risk management, including improving the development of preventive methods in the professional selection, help modernize accounting, and analyze the risk of occupational diseases. At the same time due to easier access to medically important data information and communication technologies provide new forms of interaction between participants in the monitoring process. The transition to a fully connected medical IoT environment will significantly reduce data entry errors and, together with IoT security tools, increase its cyber resilience. In this context, reliable communication and interaction, along with mobility and the Internet of Things, are important tools for developing effective systems for monitoring the state of human health in the working area. The developed system will further provide a highly qualified medical service and reduce the cost of diagnostic, treatment, and prevention activities. This framework fully satisfies the requirements of the world community to improve the system of medicine and labor protection, in particular those workers who perform high-risk work, and work where there is a need for professional selection.

References 1. Litovchenko, O., Zub, K., Zavgorodnii, I., Myasoedov, V.: Society as an essential environmental factor. Inter Collegas 5(4), 155–160 (2018) 2. Perova, I., Litovchenko, O., Bodvanskiy, Y., Brazhnykova, Y., Zavgorodnii, I., Mulesa, P.: Medical data-stream mining in the area of electromagnetic radiation and low temperature influence on biological objects. In: 2018 IEEE Second International Conference on Data Stream Mining & Processing (DSMP) Proceedings, pp. 3–6. IEEE (2018) 3. Yeremenko, O., Lemeshko, O., Persikov, A.: Secure routing in reliable networks: proactive and reactive approach. In: Shakhovska, N., Stepashko, V. (eds.) CSIT 2017. AISC, vol. 689, pp. 631–655. Springer, Cham (2018). https://doi.org/10.1007/978-3-319-70581-1_44

Framework for Developing a System for Monitoring Human Health

409

4. Yeremenko, O.: Enhanced flow-based model of multipath routing with overlapping by nodes paths. In: 2015 Second International Scientific-Practical Conference Problems of Infocommunications Science and Technology (PIC S&T) Proceedings, pp. 42–45. IEEE (2015). https:// doi.org/10.1109/INFOCOMMST.2015.7357264 5. Zavgorodnij, I., Thielmann, B., Kapustnik, W., Batschinskij, R., Batschinskaja, J., Böckelmann, I.: Toxizität von Methyl-tert-butylether auf innere Organe von Versuchstieren unter Kältebedingungen. Zentralblatt für Arbeitsmedizin, Arbeitsschutz und Ergonomie 69(5), 290–300 (2019) 6. Piˇnosová, M., Andrejiová, M., Lumnitzer, E.: Synergistic effect of risk factors and work environmental quality. Calitatea 19(165), 154–159 (2018) 7. Chernyuk, V.I., Nazarenko, V.I., Apykhtin, K.O., Chui, T.S., Kudievskiy, Y.V.: Physiological and hygienic evaluation of physical factors and intensity at work places of flight operators and engineering staff in air traffic management in Ukraine. Ukrainian J. Occup. Health 1(29), 26–33 (2012) 8. Tan, S., Wang, H., Peng, R.: A review on combined biological effects of microwave and other physical or chemical agents. Int. J. Radiat. Res. 16(2), 139–153 (2018) 9. Abdi, S., Dorranian, D., Naderi, G.A., Razavi, A.E.: Changes in physicochemical charachteristics of human low density lipoprotein nano-particles by electromagnetic field exposure. Stud. Univ. Babes-Bolyai, Chem. 61(1), 185–197 (2016) 10. Sheikh, M., et al.: Individual and combined effects of environmental risk factors for esophageal cancer based on results from the golestan cohort study. Gastroenterology 156(5), 1416–1427 (2019) 11. Ben-Israel, D., et al.: The impact of machine learning on patient care: a systematic review. Artif. Intell. Med. 103, 101785 (2020). https://doi.org/10.1016/j.artmed.2019.101785 12. Aqueveque, P., Gutierrez, C., Rodríguez, F.S., Pino, E.J., Morales, A.S., Wiechmann, E.P.: Monitoring physiological variables of mining workers at high altitude. IEEE Trans. Ind. Appl. 53(3), 2628–2634 (2017) 13. Pasha, M., Shah, S.M.W.: Framework for E-Health systems in IoT-based environments. Wirel. Commun. Mob. Comput. 6183732, 1–11 (2018). https://doi.org/10.1155/2018/6183732 14. Gusev, M., Ristov, S., Prodan, R., Dzanko, M., Bilic, I.: Resilient IoT eHealth solutions in case of disasters. In: 2017 9th International Workshop on Resilient Networks Design and Modeling (RNDM) Proceedings, pp. 1–7. IEEE (2017). https://doi.org/10.1109/RNDM.2017.8093024 15. Perova, I., Bodyanskiy, Y.: Adaptive human machine interaction approach for feature selection-extraction task in medical data mining. Int. J. Comput. 17(2), 113–119 (2018) 16. Perova, I., Brazhnykova, Y., Bodyanskiy, Y., Mulesa, P.: Neural Network for Online Principal Component Analysis in Medical Data Mining Tasks. In: 2018 IEEE First International Conference on System Analysis & Intelligent Computing (SAIC) Proceedings, pp. 1–5. IEEE (2018) 17. Shakhovska, N., Medykovsky, M., Stakhiv, P.: Application of algorithms of classification for uncertainty reduction. Przeglad Elektrotechniczny 89(4), 284–286 (2013) 18. Shakhovska, N., Vovk, O., Kryvenchuk, Y.: Uncertainty reduction in big data catalogue for information product quality evaluation. Eastern-Eur. J. Enterp. Technol. 1(2 (91)), 12–20 (2018). https://doi.org/10.15587/1729-4061.2018.123064 19. Shirazi, S.N., Gouglidis, A., Farshad, A., Hutchison, D.: The extended cloud: Review and analysis of mobile edge computing and fog from a security and resilience perspective. IEEE J. Sel. Areas Commun. 35(11), 2586–2595 (2017). https://doi.org/10.1109/JSAC.2017.276 0478 20. Prokhorenko, V., Babar, M.A.: Architectural resilience in cloud, fog and edge systems: a survey. IEEE Access 8, 28078–28095 (2020). https://doi.org/10.1109/ACCESS.2020.297 1007

410

O. Yeremenko et al.

21. Arooj, M., Asif, M.: Sensing as a service: vision, practices and architecture. Int. J. Inf. Eng. Electr. Bus. 11(6), 37–43 (2019). https://doi.org/10.5815/ijieeb.2019.06.06 22. Kavitha, S., A.Alphonse, P.J.A.: A hybrid cryptosystem to enhance security in IoT health care system. Int. J. Wireless Microwave Technol. 9(1), 1–10 (2019). https://doi.org/10.5815/ ijwmt.2019.01.01 23. Kumar, V., Laghari, A.A., Karim, S., Shakir, M., Brohi, A.A.: Comparison of fog computing & cloud computing. Int. J. Math. Sci. Comput. (IJMSC) 5(1), 31–41 (2019). https://doi.org/ 10.5815/ijmsc.2019.01.03 24. Abdallah, H.M., Taha, A., Selim, M.M.: Cloud-based framework for efficient storage of unstructured patient health records. Int. J. Comput. Network Inf. Secur. 11(6), 10–21 (2019). https://doi.org/10.5815/ijcnis.2019.06.02 25. Kame, S.O.M., Elhamayed, S.A.: Mitigating the impact of IoT routing attacks on power consumption in IoT healthcare environment using convolutional neural network. Int. J. Comput. Network Inf. Secur. 12(4), 11–29 (2020). https://doi.org/10.5815/ijcnis.2020.04.02

Chinese College Students’ L2 Motivational Self System in the Context of Artificial Intelligence Jie Ma1(B) and Pan Dong2 1 Wuhan Technology and Business University, Wuhan 430065, China 2 South-Central University for Nationalities, Wuhan 430074, China

Abstract. Based on the theory of L2 Motivational Self System, the English Learning Motivation Questionnaire is used in this research to investigate the factors, characteristics and specific content of 212 Chinese college students’ L2 Motivational Self System in the context of artificial intelligence-based language learning. The findings of this research are the following: L2 Motivational Self System contains three factors Ideal L2 Self, Ought-to L2 Self, and L2 Learning Experience; Ideal L2 Self is dominant in the students’ L2 Motivational Self System; Ideal L2 Self contains two factors: Ideal L2 Communicator and Ideal L2 User; Ideal L2 User is significantly higher than Ideal L2 Communicator; artificial intelligencebased language learning promotes learners’ communicative vision, translation and written language communication expectations, and initiative for language learning. The research further proposes to rationally guide the formation of learners’ Ideal L2 Self, and use the oral training, machine translation and adaptive learning system provided by artificial intelligence technology to generate and maintain college students’ language learning motivation. Keywords: L2 motivational self system · Ideal L2 self · Artificial intelligence-based language learning

1 Introduction Since the 1950s, when the concept of artificial intelligence was proposed, artificial intelligence has been increasingly used in various fields such as economy, society, and education. In the field of language education, the development of artificial intelligence has greatly boosted language learning, especially second language learning [1]. This research makes no distinction between second language learning and foreign language learning, and both refer to learning a language other than the mother tongue [2]. Artificial intelligence technology, on the one hand, enhances learners’ interest in English learning, thereby enhancing their motivation to learn English; on the other hand, artificial intelligence technology reduces the expected value of college students’ language learning, thereby affecting negatively on their English learning motivation [3]. Motivation, one of the important individual factors in second language learning, has a direct impact on second language learning [4]. Therefore, the study of second language learning motivation in the age of artificial intelligence helps researchers to acquire language learners’ © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 411–422, 2021. https://doi.org/10.1007/978-3-030-80472-5_34

412

J. Ma and P. Dong

current motivational situation, and is beneficial for teachers to use artificial intelligence technology to promote the strengthening and maintenance of learners’ second language learning motivation [5]. The aims of this research are to investigate the overall situation of college students’ second language learning motivation and to make suggestions for teachers to instruct and coordinate learners’ motivation in their teaching activities by referring to the data collected in the research [6]. In this research, the language learning assisted by artificial intelligence technology is called artificial intelligence-based language learning. The devices of artificial intelligence-based language learning mainly include three categories: robots, professional software, and online learning platforms. Robots, such as iFLYTEK translators, chat robots, etc.; professional software, such as Google Translate, Kingsoft, etc.; network teaching platforms, such as fanya.chaoxing.com, China University MOOC, etc. In this survey, artificial intelligence-based language learning refers to language learning with the above three types of artificial intelligence devices.

2 L2 Motivational Self System and Ideal L2 Self The theory of L2 Motivational Self System is the latest perspective in motivation research so far [7]. The system includes Ideal L2 Self, Ought-to L2 Self, and L2 Learning Experience. The biggest contribution of this theory is that it clarifies that Ideal L2 Self is more in line with motivational research in the EFL learning environment [8]. Ideal L2 Self in this research refers to a kind of emotional and cognitive construction of the learner’s future self in foreign language learning (Fig. 1).

Ideal L2 Self

L2 Motivational Self

Ought-to L2 Self

L2 Learning Experience Fig.1. L2 motivational self system

Researchers have carried out related researches on the validity of L2 Motivational Self System in the foreign language learning environment, its contribution to motivation and its measurement tools. Sun Lei and Lv Zhongshe have taken English majors and non-English undergraduates and postgraduates as the research objects, verifying the importance and possibility of Ideal L2 Self, especially the contribution of “possibility” on motivation [9]. Zhan Xianjun, taking Chinese college students and EFL undergraduates in Sino-US cooperative classes as the research objects, has conducted a multi-group verification of Ideal L2 Self in the Chinese context and pointed out that the Ideal L2 Self is a multi-dimensional structure [10]. Wei Yaoyu and Fan Weiwei compiled a more efficient and reliable “Chinese College Students L2 Motivational Self System through a

Chinese College Students’ L2 Motivational Self System in the Context

413

questionnaire survey of 1331 non-English major college students [11]. Sun Yunmei and Li Zan have used meta-analysis to explore the relationship between the dimensions of L2 Motivational Self System and the learners’ expected effort and language achievement in language acquisition, and the moderating effect of related variables on this relationship [12]. However, none of the above studies has been conducted on college students in the context of artificial intelligence. This research, taking college students as the research objects and clarifying the composition, characteristics of and the specific content of the L2 Motivational Self System of these objects in language learning under the background of artificial intelligence, expands the verification group of the researches on L2 Motivational Self System, and improves the currency of motivation research.

3 Research Design 3.1 Research Questions This research adopts a mixed design combining quantitative and qualitative methods to explore the following issues: 1) In artificial intelligence-based language learning, what are the constituent factors of college students’ L2 Motivational Self System? 2) In artificial intelligence-based language learning, what are the characteristics of college students’ L2 Motivational Self System? 3) In language learning based on artificial intelligence, what are the specific content of college students’ Ideal L2 Self? 3.2 Participants Participants in this study are 212 students from a ministry-affiliated college and a provincial college in Wuhan, Hubei Province, with an average age of 19, from ideological and political, sociology, marketing, e-commerce, logistics, online new media, environmental and biological engineering, pharmaceutical and applied chemistry, etc. The average time for participants to use the three types of artificial intelligence language learning devices: robots, professional software and online learning platforms is 2.33 years. 3.3 Instrument The instruments of this research include questionnaires and an interview question. The questionnaire English Learning Motivation Questionnaire consists of two parts: the first part is background information, including the age, gender and time of the students’ use of artificial intelligence-based language learning devices; the second part is the L2 Motivational Self System Scale, a total of 29 items. This scale is compiled with reference to Wei Yaoyu and Fan Weiwei’s [11]. Chinese College Students’ L2 Motivational Self System Scale and Ought-to L2 self Scale, combined with the English Learning Questionnaire by Liu Zhen et al. [13]. The questionnaire uses a Likert 5-point measurement method. The interview question is compiled based on the preliminary analysis results of the quantitative data of the questionnaire. It contains one question, distributed through the

414

J. Ma and P. Dong

online learning platform. Participants are required to describe as much as possible in Chinese or English language in order to collect the specific content of college students’ Ideal L2 Self. 3.4 Data Collection and Analysis In April 2020, the two authors of this article have distributed 233 questionnaires and recovered 212 valid questionnaires. The effective recovery rate of the questionnaire is 90.99%. After the questionnaire survey, we collected qualitative data by distributing an interview question (n = 47) to some students on the online learning platform. Students can answer by voice or text in Chinese or English. The voice is converted into text by the first author of this article. The text data describing Ideal L2 Self contains 4852 Chinese characters and 969 English words. Quantitative data processing uses SPSS 21.0, and qualitative data uses content analysis. The research consists of 3 steps: 1) In exploratory factor analysis, principal component analysis is adopted, and according to the maximum variance method the internal factors of college students’ L2 Motivational Self System are explored; paired sample t-test is used to verify the significance of the difference in internal factors of Ideal L2 Self. 2) In exploratory factor analysis, principal component analysis is adopted, and according to the maximum variance method the internal factor structure of the Ideal L2 Self of college students is explored; paired sample t-test is used to verify the significance of the difference in internal factors of Ideal L2 Self. 3) Content analysis is used to classify and analyze the specific content of the Ideal L2 Self. Before the analysis, a coding table has been established based on the factors of the Ideal L2 Self, and then the qualitative data has been coded with the coding table, and the coding table has been revised in the coding process.

4 Results and Discussion 4.1 Factors of L2 Motivational Self System The first author of this article verified the structural validity and reliability of the English Learning Motivation Questionnaire. The KMO value of the scale data is .902, and the associated probability value of the Bartlett sphere test is .000, indicating that the data is suitable for factor analysis. Exploratory factor analysis adopts principal component analysis method, and performs factor analysis on the data according to the maximum variance method. The load value of each item is above .40, and there is no item with a load above .40 on both factors, which is classified into three factors, explaining 56.996% of the total variance. The internal reliability analysis of the scale shows that the overall Cronbach’s alpha coefficient of the scale is .936, and the Cronbach’s alpha coefficients of the subscales are .933, .927 and .796 in order. The questionnaire structure is reliable. The three factors are named Ideal L2 Self, L2 Learning Experience, and Ought-to L2 Self according to variance contribution rate (see Table 1).

Chinese College Students’ L2 Motivational Self System in the Context

415

Table 1. Descriptive statistics of the L2 motivational self system scale Scale

Factor

Mean

Std. Deviation

Cronbach’s alpha

L2 Motivational Self System

Ideal L2 Self

3.0500

.92026

.933

L2 Learning Experience

2.8296

.92029

.927

Ought-to L2 Self

2.7586

.88627

.796

4.2 Characteristics of L2 Motivational Self System 4.2.1 Ideal L2 Self: The Key Factor of L2 Motivational Self System From the means in Table 1, Ideal L2 Self is dominant in the participants’ L2 Motivational Self System. In order to verify the significance of the difference within Ideal L2 Self, L2 Learning Experience, and Ought-to L2 Self, the first author of this article has used a paired sample t test to test the subjects’ difference within Ideal L2 Self, L2 Learning Experience, and Ought-to L2 Self. Comparative analyses have been carried out, and the results have showed that Ideal L2 Self is significantly higher than L2 Learning Experience, and Ought-to L2 Self (see Tables 2 and 3). Table 2. The difference of experience level between ideal L2 self and ought-to L2 self Ideal L2 Self Experience Level

Ought-to L2 Self

M

SD

M

SD

3.05

.92

2.76

.89

MD

t(211)

.29

3.996*

* p < 0.05

Table 3. The difference of experience level between ideal L2 self and L2 learning experience Ideal L2 Self Experience Level

L2 Learning Experience

M

SD

M

SD

3.05

.92

2.83

.92

MD

t(211)

.22

3.623*

* p < 0.05

These findings are consistent with the research results of Zhan Xianjun, indicating that the Ideal L2 Self is the core concept and key component of L2 Motivational Self System [10]. Ideal L2 Self includes the emotional imagination and psychological cognition of the future self under the foreign language learning situation. These “tangible images related to the acquisition goal” enable the ideal second language self to perform the emotional and motivational functions. Learners generate motivation in the learning process, and strive to narrow the gap between the real self and the ideal self, so that learning motivation

416

J. Ma and P. Dong

and self-concept are closely related and become the source of motivation [9]. Therefore, Ideal L2 Self has become the dominant factor of language learners’ L2 Motivational Self System. 4.2.2 Factors of Ideal L2 Self: Ideal L2 Communicator and Ideal L2 User In order to explore the constituent factors of Ideal L2 Self, the first author of this article has conducted an exploratory factor analysis on the data of Ideal L2 Self Scale in the questionnaire. The first round of results has showed that KMO = .915 and Bartlett’s spherical shape is significant (p = .000), indicating that the data is suitable for factor analysis. Principal component analysis is used for exploratory factor analysis, and the item items is to be deleted if the following analysis results are obtained: 1) The maximum load value is less than 0.4; 2) the factor characteristic value is less than 1; 3) more than one factor load value is greater than 0.4; 4) the factor item is less than 3. 8 items are deleted, and the remaining 7 items are classified into 2 factors, explaining 66.420% of the total variance. The internal reliability analysis of the scale shows that the Cronbach’s alpha coefficient of the overall scale is .844, and the Cronbach’s alpha coefficients of the two factors are .797 and .792. The questionnaire structure is reliable. The two factors are named Ideal L2 User and Ideal L2 Communicator in descending order of variance contribution rate (see Table 4). Table 4. Descriptive statistics of the Ideal L2 self scale Scale

Factor

Mean

Std. Deviation

Cronbach’s alpha

Ideal L2 Self

Ideal L2 Communicator

2.6439

.99044

.797

Ideal L2 User

3.6604

1.03948

.792

The results of the two factors of Ideal L2 User and Ideal L2 Communicator are consistent with the research results of Wei Yaoyu and Fan Weiwei [11]. The factor Ideal L2 Communicator is determined by the primary function of language is the communicative function. The self-image and future states related to successful English communication, such as the “successful L2 speaker”, have become common findings of previous related studies [11, 14, 15]. The generation of the factor Ideal L2 User is the product of the foreign language learning situation. Due to limited communication opportunities with native speakers, media such as the Internet and television have become the main channels for foreign language learners to learn English and understand international trends. The language materials provided by the media enrich the learning content of foreign language learners, and at the same time arouse and attract the interest of the learners, and learners have an expectation of effective use of the media to obtain these learning content. This interactive relationship between media use and English learning corresponds to the discovery of “English media-oriented motivation” by Clement et al. [16]. The generation of this self-image reflects the important role of the media in the English learning of foreign language learners.

Chinese College Students’ L2 Motivational Self System in the Context

417

4.2.3 Characteristics of Ideal L2 Self: Ideal L2 User Significantly Higher Than Ideal L2 Communicator In order to clarify the characteristics of the participants’ Ideal L2 Self and reveal the difference between its two factors, the first author of this article has conducted a comparative analysis of these two factors through a paired sample t test. The results show that the Ideal L2 User is significantly higher than Ideal L2 Communicator (see Table 5). In addition, according to Oxford and Burry-Stock’s classification standard of Likert’s 5-point scale, the average value is equal to or higher than 3.5 for high frequency use, and the average value is introduced. The level between 2.5 and 3.4 is moderate, and the average value is equal to or lower than 2.4 is low-frequency use [17]. Table 5 shows that the participants’ Ideal L2 User reach the high-frequency experience level. Ideal L2 User reaches the high-frequency experience level because Ideal L2 User contains “sing English songs” and “easily watch English programs”, compared with the “ideal second language communicator” contains “speech or debate”, “work in English” and other content are one-direction use of English, which is easier and more likely to be realized than two-direction interactive communication. Possibility plays a more important role in promoting learning motivation, the higher the possibility of achieving the goal, the easier it is for people to stimulate higher motivation [8]. Table 5. The difference of experience level between ideal L2 user and ideal L2 communicator Ideal L2 User Experience Level

Ideal L2 Communicator

M

SD

M

SD

3.66

1.04

2.64

.99

MD

t (211)

1.02

15.676*

* p < 0.05

4.3 Specific Content of Ideal L2 Self In questionnaire survey, the constituent factors of the participants’ Ideal L2 Self include Ideal L2 Communicator and Ideal L2 User. Ideal L2 Learner is a new finding different from the results of the questionnaire survey. 4.3.1 Ideal L2 Communicator: Enhancement of Oral Communicative Vision and Learning Confidence by AI Technology Artificial intelligence-based language learning devices provide ubiquitous, flexible and intelligent oral training, which enhances learners’ communicative vision and learning confidence. Student Y describes that she can “download a software for practicing oral English on my mobile phone, which contains countless practice videos and tutorials and artificial intelligence can supervise my study time and rest time”. From her description, it can be found that student Y’s strengthening of the vision of oral communication with the help of artificial intelligence technology is mainly due to two factors: one is the rich variety of oral training materials and the convenience of downloading; the other is that

418

J. Ma and P. Dong

artificial intelligence technology provides a training plan, which has a supervising effect on oral training. Language learning assisted by AI technology improves students’ oral and learning initiative and participation. Repeated imitation, self-correction, and use of big data to record learning processes and behaviors to achieve intelligent and standardized evaluation, these all contribute to the sustainable learning interest of learners [18]. Student Z has mentioned that artificial intelligence provides a partner for oral practice, which improves the learning efficiency of oral training. Artificial intelligence can “reduce the barriers of face-to-face communication” and she can “practice oral English by network”. At present, with the rapid development of technologies such as speech recognition, speech synthesis, natural speech understanding, and knowledge base retrieval, artificial intelligence-based language learning devices have begun to have intelligent language interaction functions. AI-driven man-machine dialogue technology has always been given high hopes in the language learning process, especially in the dialogue practice and context simulation of language learners. First of all, artificial intelligence technology can simulate real people to act as linguistic companions to communicate with learners and practice. Secondly, learners can easily construct a language environment through dialogue with the machine, and can even easily switch languages to achieve efficiency and functions that cannot be achieved by real learning partners. Thirdly, while realizing multi-functions, human-computer dialogue uses intelligent software to act as a real teacher, which greatly reduces the cost of language learners [19]. 4.3.2 Ideal L2 User: Enhancement of Translation and Written Language Communication Expectation by AI Technology With the assistance of artificial intelligence, learners’ barriers to the translation of written language are greatly reduced, and the expectation of written language communication through translation is strengthened. Student X has strong expectations for translation and he has described that artificial intelligence translation is characterized by automation and speed, such as, “correcting grammar and spelling mistakes automatically, saving time by voice typing and scanning the texts”. It can be seen that, in the context of artificial intelligence technology, online translation has the characteristics of high efficiency and convenience, which can save a lot of time, material and manpower, etc., as Zhang Yanlu pointed out, “the emergence of cloud service-based English translation software and platform can help students solve many difficulties in translation, enhance students’ confidence in learning to a certain extent, and stimulate in-depth learning interest” [20]. Machine translation gradually reduces the difficulty of translation and promotes learners’ expectations of written language expression and communication. Student L has hoped that his works can be read by readers from other countries, expressing the expectation of communicating through his own writing and being translated. He has described that translation assisted by artificial intelligence can “translate my own work accurately, … and the text content is original.” Driven by the increasingly mature artificial intelligence technology, machine translation has achieved multiple paradigm shifts, and gratifying results in translation speed, translation quality, language coverage, and knowledge acquisition [21]. The accuracy and diversity of intelligent translation are

Chinese College Students’ L2 Motivational Self System in the Context

419

improved, and its performance is close to or even comparable to that of manual translation. Intelligent language learning tools that do not rely on manual intervention are coming to us [18]. 4.3.3 Ideal L2 Learner: Promotion of Initiative in Language Learning by Adaptive Learning System of AI Technology Through the artificial intelligence teaching platform, student F’s learning effects are consolidated and learning content is expanded. He has described that he can “successfully answer questions and get high scores in the English test” because he can “use the resources on the Internet, review and consolidate [what he has learned] on the Internet again”. This shows that the design of the intelligent foreign language teaching platform takes text-deep-mining technology and visualization technology application as the core, realizes the automation, data and visualization of the whole teaching process, effectively stimulates students’ interest in learning, and fully activates the enthusiasm and initiative of learning [22]. It is worth mentioning that language learning with artificial intelligence intervention provides adaptive learning, which refers to the process by which learners acquire knowledge through an adaptive learning system supported by computer technology. The adaptive learning system participates in the entire learning process of learners, provides personalized learning services, and plays a role in learning plan formulation, learning strategy organization, knowledge composition design, and learning effect evaluation. Artificial intelligence has become the key to the formation of adaptive learning support systems. The adaptive learning system stimulates learning interest and enhances the self-consciousness and initiative of learning through psychological analysis of learning [23].

5 Conclusion and Implications This research is based on the theory of L2 Motivational Self System and uses the English Learning Motivation Questionnaire to investigate the factors and characteristics of students’ L2 Motivational Self System and the status, factors and characteristics of Ideal L2 Self of 212 college students. The research has found the following results: L2 Motivational Self System contains three factors Ideal L2 Self, Ought-to L2 Self, and L2 Learning Experience; Ideal L2 Self is dominant in participants’ L2 Motivational Self System; Ideal L2 Self contains two factors: Ideal L2 Communicator and Ideal L2 User; Ideal L2 User is significantly higher than Ideal L2 Communicator, and the participants’ Ideal L2 User has reached the high-frequency experience level. Through the analysis of the description texts of 47 subjects’ Ideal L2 Self, Ideal L2 Learner is a new finding. To be specific, artificial intelligence-based language learning promotes learners’ oral communicative vision and learning confidence, translation and written language communication expectations, and enhances learning initiative (Fig. 2). Based on the above findings, the author suggests that in future English teaching, teachers can guide students to maintain a positive Ideal L2 Self from the following three aspects.

420

J. Ma and P. Dong

Ideal L2 Communicator

Ideal L2 User

Ideal L2 Self

Ideal L2 Learner Fig.2. Ideal L2 self system

1) Before the start of foreign language courses, teachers should help students create learning goals and visions, help students clarify the specific content of Ideal L2 Self, and maintain students’ English learning motivation. When setting goals, teachers need to follow three principles: (1) the goals are challenging; (2) the goals are possible; (3) long-term and short-term goals are combined. Teachers help students create, consolidate and strengthen the vision, guide the students to transform the goal and vision of Ideal L2 Self into action, maintain the vitality of the vision, and promote Ideal L2 Self to maximize the effect because motivation is dynamic in nature and can vary from moment to moment depending on the learning context or task [24]. 2) Teachers should promote students to use the learning media and devices assisted by artificial intelligence technology more effectively, use the context simulation and machine translation provided by artificial intelligence technology to consolidate and strengthen the goals and visions of the Ideal L2 Communicator and Ideal L2 User, the guiding role of the training to promote the maintenance of oral training motivation. At the same time, corresponding reforms should be made to leaning goals and curriculum settings to effectively meet the needs of language intelligence development. 3) In teaching practice, teachers should realize that the deep integration of artificial intelligence and language learning process will become the new way of language teaching. Teachers are no longer the leaders of teaching activities, but the guides and coordinator in the individualized learning process of students. Teacher plays the role of learning expert in the adaptive learning provided by artificial intelligence, mainly engaged in teaching intervention and teaching governance, providing learning content, teaching design, participating in teaching management, and guiding learners to create a personalized learning environment. From systematic teaching service, teaching process guidance and systematic training plan formulation, artificial intelligence technology promotes learners to establish a personalized adaptive learning system so that language learning can become more effective and efficient.

Acknowledgment. This project is supported by supported by Humanities and Social Science Research Project of Hubei Provincial Department of Education (QSY17007), Special Fund of Cultivation Project for Basic Scientific Research of Central Universities (CSP17028) and Teaching Research Project of South-Central University for Nationalities (Jyx16033).

Chinese College Students’ L2 Motivational Self System in the Context

421

References 1. Bingyan, H., Yumei, L.: A study of cognitive basis of AI language and its new implications for SLA[J]. Tefle 2, 78–83 (2019). (in Chinese) 2. Chen Lin, X., Yan: The exclusive search engine designed for personalized second language Chinese learner. Int. J. Educ. Manage. Eng. 2(11), 8–13 (2012) 3. Jifang, F.: The impact of artificial intelligence on college English education. Heilongjiang Educ. 7, 85–86 (2020). (in Chinese) 4. Haichao, C.: Directed motivational current: a new development in L2 motivational theory. Mod. Foreign Lang. 39(5), 704–712 (2016). (in Chinese) 5. Ying, W.: The effect of connecting reading and writing in foreign language teaching in the computer mediated communication environment. Int. J. Educ. Manage. Eng. 2(10), 25–29 (2012) 6. Gilakjani, A.P.: The significant role of multimedia in motivating EFL learners’ interest in English language learning. Int. J. Mod. Educ. Comput. Sci. 4(4), 57–66 (2012) 7. Dörnyei, Z.: The Psychology of the Language Learner: Individual Differences in Second Language Acquisition. Lawrence Erlbaum Associates, Mahwah, NJ (2015) 8. Fengge, L.: Validity analysis of the theoretical model of l2 motivational self system among various groups of English learners in China. Flltp 1, 19–23 (2015). (in Chinese) 9. Lei, S., Zhongshe, L.: Importance and possibility: the contribution of the idea L2 self to L2 learning motivation. Foreign Lang. Res. 6, 53–58 (2014). (in Chinese) 10. Xianjun, Z.: Validating ideal l2 self (IL2S) across groups in the Chinese context[J]. Foreign Lang. Teach. Res. 47(4), 573–584 (2015). (in Chinese) 11. Yaoyu, W., Weiwei, F.: Compilation of ideal l2 self and ought-to l2 self scale for Chinese college students. J. PLA Univ. Foreign Lang. 39(6), 77–85 (2016). (in Chinese) 12. Yunmei, S., Zan, L.: A meta-analysis of the relationship between l2 motivational self system and language acquisition. Mod. Foreign Lang. 43(6), 830–840 (2020). (in Chinese) 13. Zhen, L., Xiaojun, Y., Sufen, H.: A structural analysis of undergraduates’ second language self, anxiety and motivational learning behavior. Foreign Lang. World 6, 28–37+94 (2012). (in Chinese) 14. Al-Shehri, A.: Motivation and vision: the relation between the ideal l2 self imagination and visual style. Motivation, Language Identity and the L2 Self, 164–171 (2009) 15. Csizer, K., Kormos, J.: Learning experiences, selves and motivated learning behaviour: a comparative analysis of structural models for hungarian secondary and university learners of English. Motivation, Language Identity and the L2 Self, 98–119 (2009) 16. Clément, R., Dörnyei, Z., Noels, K.A.: Motivation, self-confidence, and group cohesion in foreign language classroom[J]. Lang. Learn. 44(3), 417–448 (1994) 17. Oxford, R.L., Burry-Stock, J.A.: Assessing the use of language strategies worldwide with the ESL/EFL version of the strategy inventory for language learning[J]. System 23(2), 1–23 (1995) 18. Tian, N.: Application and challenge of artificial intelligence technology in language learning. Mod. Inf. Technol. 3(19), 109–110 (2019). (in Chinese) 19. Tian, N.: Definition, design and innovation of intelligent language learning software in the era of artificial intelligence[J]. Electr. World 21(11), 45–46 (2019). (in Chinese) 20. Yanlu, Z.: The study on promotion of artificial intelligence to the reform of college English teaching[J]. Educ. Modernization 7, 130–131 (2019). (in Chinese) 21. Hanjie, L., Haiqing, C.: Philosophical reflection on the technical dilemma of machine translation. J. Dalian Univ. Tech. 41(6), 122–128 (2020). (in Chinese) 22. Kaibao, H., Xujun, T.: MTI talent training in the context of language intelligence: challenges. Countermeasures Prospects, Foreign Lang. World 2, 59–64 (2020). (in Chinese)

422

J. Ma and P. Dong

23. Jianseng, M.: Adaptive learning system: implementation logic of blended teaching in the background of artificial intelligence. Educ. Reaching Forum 43, 321–323 (2020). (in Chinese) 24. Abbas, P.G., Leong, L., Narjes, B.S.: A study on the role of motivation in foreign language learning and teaching. Int. J. Mod. Educ. Comput. Sci. (7), 9–16 (2012)

Encoding Sememe and Context Information into Sentence Representation for Implicit Sentiment Analysis Qizhi Qiu and Junan Qiu(B) School of Computer Science and Technology, Wuhan University of Technology, Wuhan 430063, China [email protected]

Abstract. Sentiment analysis is one of the highly popular fields in natural language processing. Most existing researches pay more attention to the identification of explicit sentiments than that of the implicit sentiments. In this paper, we develop our analysis on implicit sentiment for Chinese text from the view of semantic information and technology. In order to enrich the semantic information of every single sentence, we propose a two-level semantic model to represent each sentence, namely, context information at the sentence level and sememe information of HowNet at the word level. On the other hand, from the view of technology, since CNN-LSTM model has achieved excellent performance on sentiment analysis, the CNN-LSTM is adopted to extract local information and long-term contextual information of the enriched sentence and the sentence representation is obtained. We conduct experiments on the SMP2019-ECISA dataset. The results show that our method outperforms other methods and realizes an improvement of 5.6% macro-average F1 score in implicit sentiment analysis. Keywords: Implicit sentiment analysis · Sememe · Context

1 Introduction Sentiment analysis has become a popular research topic due to its vast potential for applications, such as public opinion analysis and review analysis. Sentiments are, by nature, subjective because they regard people’s subjective views, appraisals, evaluations, and feelings [1]. Sentiment analysis is to identify the sentiment polarities or emotions in the text, regardless of whether the sentiment is explicit or implicit. The rapid development of deep learning has yielded significant fruit on explicit sentiment analysis [2]. Existing researches have applied CNN, LSTM model for sentiment classification and got satisfactory results [3]. However, the analysis of implicit sentiment, which contains no sentiment clues is a more challenging task [4, 5]. Unlike explicit sentiment analysis, the knowledge base and context information play a vital role during implicit sentiment analysis. There have been recent researches for implicit sentiment analysis based not just on commonsense knowledge bases, but also © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 423–433, 2021. https://doi.org/10.1007/978-3-030-80472-5_35

424

Q. Qiu and J. Qiu

the characteristics and the triggers of implicit sentiment. Some researchers applied ConceptNet into implicit sentiment analysis and proved the effectiveness [6, 7], While others analyzed the reasons for triggering sentiments and observed that context information is useful for implicit sentiment analysis [8, 9]. Implicit sentiment analysis requires strategies that combine the sentence with a knowledge base rather than bag-of-words and etc. [10]. In order to capture the limited information of single sentence during implicit sentiment analysis, this paper firstly adopts a novel encoding strategy to represent the target sentence, and propose a two-level semantic feature model, where we combine the semantic information of every notional word and context information of the sentence during implicit sentiment analysis. In our solution, HowNet provides the semantic information for word in form of sememe, which refers to the smallest semantic unit and contains more paraphrase of words and concepts. And then we take advantage of CNN-LSTM to extract local information and long-term contextual information. Therefore, we use the representation of sentence obtained by CNN-LSTM to classify implicit sentiment sentence.

2 Related Work 2.1 Representation Learning and Explicit Sentiment Analysis Applying deep neural networks to sentiment analysis has become popular. Within the natural language processing field, many tasks of deep learning methods have involved representation learning [11]. Existing approaches to learning sentence representation for sentiment analysis fall into two categories: (1) improve or combine neural network models; (2) encoding external sentiment resources such as sentiment lexicon. Methods in the former category usually require little external domain knowledge and reach satisfactory results in sentiment analysis. Kim et al. applied CNN for a series of sentence representation learning tasks and got competitive performance on sentiment classification [12]. Wang et al. worked on twitter text and utilized LSTM for sentiment classification [5]. Zhou et al. indicated that the CNN-LSTM outperformed both CNN and LSTM [4]. Wang et al. presented a tree-structured regional CNN-LSTM model for sentiment analysis. They employed a tree-structured region division strategy to identify different depths of a parser tree and incorporated the structural information to improve performance [13]. In contrast, methods fall into the other category need to incorporate external knowledge and other features information into word vector and representations. Mumtaz focused on sentiment lexicon for sentiment analysis [14]. Ye et al. encoded sentiment knowledge into pre-trained word vectors, which effectively improved the accuracy of sentiment analysis [15]. Xu et al. proposed the Emo2vec model, which trained on six different sentiment-related tasks to encode sentiment information into the vectors, and performed better than existing affect-related representation [16]. Li et al. incorporated POS (Part of Speech) and lexicon information during the word representation training process and achieved the best performance on the evaluation dataset in NLPCC 2013 [17].

Encoding Sememe and Context Information into Sentence Representation

425

2.2 Implicit Sentiment Analysis Most of the research performed in the field of sentiment analysis has aimed at detecting explicit expressions of sentiment. However, sometimes sentiment is not expressed by the use of emotion-bearing words [6]. When there are no explicit sentiment clues in the sentences, the existing explicit sentiment analysis methods will no longer be applicable. To address implicit sentiment analysis, some researchers have focused on commonsense knowledge base construction. Balahur et al. built a commonsense knowledge base—EmotiNet with the concept of affective value based on ontology method, and proposed a method for detecting sentiment based on EmotiNet, by this way sentiment could be automatically detected without obvious emotional clues [6]. Yang et al. built a Chinese sentiment knowledge base based on semantic resource—ConceptNet, and verified their solution on public datasets [7]. Some researchers have paid attention to the characteristics and the triggers of implicit sentiment. Liao et al. analyzed the characteristics of fact-implied implicit sentiment sentences and considered that fact-implied implicit sentiment was usually affected by sentiment target, context and the structure of the sentence [9]. The investigation of Chen et al. showed that implicit sentiment opinions had the same sentiment polarity with its context [8]. Balazs et al. found that hashtags and emojis provided rich sentiment information, while POS acted as an additional input to improve model performance [18]. As mentioned, neural network models, such as CNN-LSTM, have yielded significant results in representation learning. However, as for the implicit sentiment sentence, it is limited to learn sentence representation by only neural network models. We studied semantic features from two levels: (1) word-level, most implicit sentiment analysis constructed commonsense knowledge based on ConceptNet. ConceptNet is generated automatically from the English sentences of the Open Mind Common Sense (OMCS) corpus, while the set of sememe in HowNet is established on meticulous examination of about 6000 Chinese characters. Compared to ConceptNet, HowNet contains more paraphrase of words and is more suitable for processing Chinese text. Niu et al. proposed a method of utilizing sememe information to improve word representation learning. The experiment is performed on word similarity and word analogy tasks to verify the effectiveness of the proposed method [19]. (2) sentence-level, existing researches have shown that the sentences have the same sentiment polarity with its contexts. Therefore, this paper use CNN-LSTM model to learn sentence representation by combining sememe and context information and finally predict the polarity of the implicit sentiment sentence.

3 Methodology In this section, we first describe the framework of the proposed model. Then we describe the process of obtaining sememe information from sentence and the impact of the context semantic background. 3.1 Framework of the Proposed Model Figure 1 shows the framework which fuses the sememe of HowNet and context information. It is a five-layer architecture.

426

Q. Qiu and J. Qiu Output Layer

Output

softmax

LSTM Layer

LSTM

LSTM

LSTM

CNN Layer

CNN

Embedding Layer

Embedding

LSTM

LSTM

...

...

Sememe

Sememe

Input Layer ...

...

Sentence

...

...

Context

Fig. 1. Our framework with sememe and context information

Figure 1 shows the framework with five layers: 1) The Input layer is to process all the input data, the original input data includes the sentence itself and its context information. Firstly, each sentence is split into separate words, noted as [w1 , w2 , . . . , wn ] and the sememe information of wi (i = 1, 2, …, n) is obtained through Sememe component. By the same way the context is processed in the input layer. 2) The Embedding layer is responsible to encode the semantic information and context information as vectors. In other words, the sentence itself, sememe of the words that are contained in the sentence, its context sentences and sememe of the words that are contained in context are encoded as vectors. 3) The CNN layer takes the encoded vectors as input and extracts local information with different kernel sizes. 4) The LSTM layer captures long-term dependencies of a sequence on top of the CNN. We regard the last hidden vector of LSTM as the representation of sentence and

Encoding Sememe and Context Information into Sentence Representation

427

context and put it into a softmax layer whose length is equal to the number of class labels. 5) The Output layer computes the conditional probability distribution of the sentiment polarity of the sentence. 3.2 Sememes in HowNet HowNet is a commonsense knowledge base unveiling inter-conceptual relations and inter-attribute relations of concepts as connoting in lexicons of the Chinese and their English equivalents [20]. There are two main concepts in HowNet: concept and sememe. A sememe refers to the smallest basic semantic unit that cannot be reduced further, while the concept can be represented by the relevant sememes. Their semantic relation is shown in Fig. 2.

baofu

cloth-wrapper

tool

put

wrap

word

burden

concept

duty

sememe

Fig. 2. An example of semantic relation of HowNet

As shown in Fig. 2, the Chinese word with pinyin “baofu” is a polysemy, HowNet unveils “baofu” with two concepts. The one is cloth-wrapper, which has three sememes: tool, put and wrap; the other concept of “baofu” is burden, which only includes one sememe: duty. In other words, HowNet considers the polysemy of words as a set of concepts according to the word’s definition. And each word usually appears in a specific context. In order to get the exact concept of the word, it is necessary to consider the context. Furthermore, HowNet defines every concept as a set of sememes. In this paper, we utilize lexical knowledge base OpenHowNet, it is an open sememebased lexical knowledge base which is based on well-known HowNet [21]. Firstly, each concept in HowNet is defined as a set of sememes, sememe is the smallest semantic unit of word meanings and is unique and certain, then we can use the sememe vector to represent the vector of concept. Secondly, there are one or more concepts of the word in HowNet, the vectors of concepts are denoted as conceptsV. Then, the vector of context of the word is represented as contextV. Finally, by calculating the similarity between contextV and conceptsV, the most similar concept is regarded as the meaning of the word in current context. The process is shown in Algorithm 1. After processing by Table 1, the word list [w1 , w2 , . . . , wn ] is transformed into the concept list [c1 , c2 , . . . , cn ], each concept is represented by its sememes, which is exactly the process performed in the Sememe component in the Input layer in Fig. 1.

428

Q. Qiu and J. Qiu Table 1. Process of select appropriate concept of word

3.3 Context Semantic Background As known, context is an important consideration during many NLP tasks. Current research has found out that the context semantic background is important to implicit sentiment analysis [8, 9]. However, they did not explicitly limit the size of the context. Undoubtedly the size of context influences the performance indexes of the task, such as accuracy, time-consuming and so on. The selection of the context size should depend on the specific NLP task. In this paper, we incorporate context information into implicit sentiment sentence to enrich its semantic information. However, if there are too many contexts for a certain sentence, it may diverge from the key idea of the certain sentence and it may have the opposite effect. Therefore, this paper considers the impact of different context sizes on implicit sentiment analysis and chooses the appropriate context window size through the experimental results in Sect. 4.3.2.

4 Experiments and Results 4.1 Datasets, Evaluation Index and Data Preprocessing Datasets: The experimental dataset is provided by the Evaluation of Chinese Implicit Sentiment Analysis in SMP-ECISA 2019, which is collected from Weibo, travel websites, product forums. The data is about some popular topics, such as Spring Festival Evening, haze pollution, Letv bankruptcy and so on. Table 2 shows the detailed statistics about the dataset.

Encoding Sememe and Context Information into Sentence Representation

429

Table 2. Statistics about the dataset Subset

Positive Neutral Negative

Training

3828

6989

3957

Development 1232

2553

1358

Testing

1902

979

919

Evaluation Index: SMP-ECISA 2019 adopts macro-average precision (MAP), macroaverage recall (MAR) and macro-average F1 (MAF1) as the evaluation index. Data Preprocessing: Text data from Weibo always contains specific symbols, such as #, @, etc. In order to get the usable input for the framework shown in Fig. 1. we follow the rules in Table 3 to identify and replace the special symbol. Table 3. Replacement for special symbols Original

Replacement

@users

USERNAME

//@users

ATUSERNAME

http://www.url URL

4.2 Experiment Setting Experimental Parameter Settings The framework of our proposed implicit sentiment analysis is as shown in Fig. 1, the 300 dimensional pre-trained word vectors play as the embedding layer, which is based on Weibo data provided by Li [22]. In the CNN layer, the filter windows are 3, 4 and 5 respectively with 200 feature maps each. In the LSTM layer, the hidden states of LSTM are 2 × 200-dim, and the mini-batch size is 128. Since the labels of testing set in the SMP-ECISA 2019 have not yet been published, this paper mainly uses training and development sets. We combine the training and development sets in our experiments and take 90% as training set and 10% for testing. During the training process, we take 10% of training set for validation. Comparative Experiment Settings Experiment I: Polarity Classification of Implicit Sentiment Sentences. We mainly consider the effect of sememe and context information in implicit sentiment analysis. The comparative experiment is as follows: Sentence (Sen): Only consider the sentence itself. 1) Sentence + Context (Sen + Con): Consider the sentence and its context information.

430

Q. Qiu and J. Qiu

2) Sentence + Sememe (Sen + Sem): Consider the sentence and sememe information of the sentence. 3) Sentence + Sememe + Context (Sen + Sem + Con): Consider the sentence, sememe information of the sentence, the context and sememe information of the context. Experiment II: Influence of Context Window Size. As we known, the size of the context is vital to the performance. Too many or too small contexts may have the opposite effect on the results of sentiment classification. we implement the experiments to get the proper size of context window. It is important firstly to statistic the sentence number of the document in the training set. The statistical results show that 95% of the documents don’t contain more than 12 sentences, so we set the context window size from 1 to 6. 4.3 Results and Analysis 4.3.1 Polarity Classification of Implicit Sentiment Sentences The implicit sentiment analysis requires the model to identify the sentences that contain implicit sentiments and classify their sentiment polarities. The performances of Experiment I in Sect. 4.2 are list in Table 4, where win represents the context window size. Acc is the abbreviation for accuracy, MAP, MAR and MAF1 are the macro-average scores of precisions, recall, and F1. The results in the table are the averages of 5 repeated experiments. Although deep learning technology brings about a fair performance while considering the sentence itself solely, adding context information at the sentence level and semantic information at the word level significantly improve the performance. Table 4 demonstrates that Method(Sen + Sem + Con) has the best performance among the four methods. Further study on Table 4 gets two following conclusions: Table 4. Results of polarity classification of implicit sentiment sentences Method (win = 4) Acc

MAP MAR MAF1

Sen

0.701 0.686 0.666 0.654

Sen+Con

0.721 0.689 0.669 0.668

Sen+Sem

0.726 0.707 0.688 0.688

Sen+Sem+Con

0.752 0.737 0.706 0.710

1) Deep semantic information of single word plays a more important role during the implicit sentiment analysis. Compared with Method(sen), Method(Sen + Sem) has improved by 2.6% at average on four evaluation indexes, which is also larger than that of Method(sen + con). On one hand, for implicit sentiment polarity classification, the sentences contain no explicit sentiment words that provide original sentiment clues. Most sentiments are implied in the meaning of single word. On the other hand,

Encoding Sememe and Context Information into Sentence Representation

431

unlike other knowledge bases, HowNet represents the semantic information in the form of 3-level hierarchy shown in Fig. 2, and sememe works as the foundation of the hierarchy. Employing HowNet sememe to enrich the word’s semantic information has boosted the performance of implicit sentiment analysis for Chinese text. 2) Merging sememe information and context information has greatly increased the performance. Compared with Method(sen), Method(Sen + Sem + Con) has improved by 5.0% at average on four evaluation indexes, which is also much larger than the linear addition of Method(Sen + Sem) and Method(Sen + Con). As shown in Fig. 1 the words of the context sentences are processed in the same way as the sentence itself and represented by HowNet sememe. Therefore, the current sentence is encoded by the sememe of itself and its context, and CNN-LSTM can learn more from the encoding scheme. In a word, this encoding scheme can take advantage of deep learning technology. 4.3.2 Influence of Context Window Size As mentioned above, Experiment II is implemented by different context window size in the implicit sentiment analysis. The results are shown in Table 5, win represents the context window size of the sentence. The results in the table are the average of 5 repeated experiments. Further investigation is as follows: Table 5. Results of polarity classification of implicit sentiment sentences with different context windows Method

win Acc

Sen

MAP MAR MAF1

–

0.701 0.686 0.666 0.654

Sen + Con 1

0.698 0.665 0.647 0.642

2

0.706 0.666 0.660 0.659

3

0.713 0.678 0.656 0.653

4

0.721 0.689 0.669 0.668

5

0.714 0.674 0.659 0.654

6

0.711 0.663 0.658 0.642

1) Adding context information, the performance is improved as a whole, except win = 1. The reason may be that human sentiment is very complicated, when win = 1, the context information related to the sentence is incomplete, causing an understanding error. 2) Acc and MAF1 generally first raise and then descend, which shows that simply increase the context window doesn’t benefit to the performance of implicit sentiment analysis. 3) win = 4 is the experimental result of the dataset in this paper. This dataset is mainly derived from Weibo and reviews, which are short texts with relatively concentrated

432

Q. Qiu and J. Qiu

expression of sentiment. It is very different from other genres such as literature. For other dataset, the size of win will depend on the type, length, and genre of the text in the dataset.

5 Conclusions With no explicit clues or limited information, the implicit sentiment analysis is challenging. In this paper, we propose a two-level semantic model to figure out the semantic deficiency of every single sentence. Specifically, we utilize context information to enrich semantic information at the sentence level and adopt sememe information to represent various concepts of each word at the word level. Experiments on the Chinese implicit sentiment analysis evaluation dataset demonstrates that the sememe information in HowNet can significantly improve the performance of implicit sentiment analysis, and the context window size has an impact on the implicit sentiment analysis and too small context may have the opposite effect. On the basis of this study, the follow-up research combined with external knowledge base and attention mechanism is expected to achieve new research results. Further research can focus on the use of attention mechanism to obtain appropriate emotional knowledge from large-scale external knowledge base.

References 1. Jiyao, W., Jian, L., Zhenfei, Y., et al.: BiLSTM with multi-polarity orthogonal attention for implicit sentiment analysis. Neurocomputing 383, 165–173 (2020) 2. Deqiang, Z.: A new hybrid grey neural network based on grey verhulst model and bp neural network for time series forecasting. Int. J. Inf. Technol. Comput. Sci. 9(1), 114–120 (2013) 3. Deqiang, Z.: Optimization modeling for GM(1,1) model based on BP neural network. Int. J. Inf. Technol. Comput. Sci. 2(27), 24–30 (2012) 4. Zhou, C., Sun, C., Liu, Z., et al.: A C-LSTM neural network for text classification, arXiv preprint arXiv:1511.08630 (2015) 5. Wang, X., Liu, Y., Sun, C., et al.: Predicting polarities of tweets by composing word embeddings with long short-term memory. In: Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pp. 1343–1353 (2015) 6. Balahur, A., Hermida, J.M., Montoyo, A.: Detecting implicit expressions of sentiment in text based on commonsense knowledge. In: Proceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sentiment Analysis, pp. 53–60 (2011) 7. Liang, Y., Fengqing, Z., Hongfei, L., et al.: Sentiment analysis based on emotion commonsense knowledge. J. Chin. Inf. Process. 33(6), 94–99 (2019). (in Chinese) 8. Chen, H., Chen, H.: Implicit polarity and implicit aspect recognition in opinion mining. In: Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pp. 20–25 (2016) 9. Jian, L., Suge, W., Deyu, L.: Identification of fact-implied implicit sentiment based on multilevel semantic fused representation. Knowl.-Based Syst. 165, 197–207 (2019)

Encoding Sememe and Context Information into Sentence Representation

433

10. Dos Santos, C., Gatti, M.: Deep convolutional neural networks for sentiment analysis of short texts. In: Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics, pp. 69–78 (2014) 11. Liang, M., Feng, L., Liangliang, C., Ming, H.: Simulation of crop evaportranspiration based on BP neural network model and grey relational analysis. Int. J. Inf. Technol. Comput. Sci. 2(29), 15–21 (2012) 12. Kim, Y.: Convolutional neural networks for sentence classification. In: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing, pp. 1746–1751 (2014) 13. Wang, J., Yu, L., Lai, K.R., et al.: Tree-structured regional CNN-LSTM model for dimensional sentiment analysis. IEEE/ACM Trans. Audio, Speech, Lang. Process. (28), 581–591 (2019) 14. Mumtaz, D., Ahuja, B.: A lexical approach for opinion mining in twitter. Int. J. Educ. Manage. Eng. 4, 20–29 (2016) 15. Ye, Z., Li, F., Baldwin, T.: Encoding sentiment information into word vectors for sentiment analysis. In: Proceedings of the 27th International Conference on Computational Linguistics. Santa Fe, pp. 997–1007 (2018) 16. Xu, P., Madotto, A., Wu, C., et al.: Emo2vec: learning generalized emotion representation by multi-task training. In: Proceedings of the 9th Workshop on Computational Approaches to Subjectivit, pp. 292–298 (2018) 17. Xiaojun, L., Hanxiao, S., Nannan, C., et al.: Research on sentiment analysis based on representation learning. Acta Scientiarum Naturalium Universitatis Pekinensis 55(01), 108–115 (2019). (in Chinese) 18. Balazs, J.A., Marrese-Taylor, E., Matsuo, Y.: IIIDYT at IEST 2018: implicit emotion classification with deep contextualized word representations. In: Proceedings of the 9th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis, pp. 50–56 (2018) 19. Niu, Y., Xie, R., Liu, Z., et al.: Improved word representation learning with sememes. In: Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pp. 2049–2058 (2017) 20. Fan, M., Zhang, Y., Li, J.: Word similarity computation based on HowNet. In: 2015 12th International Conference on Fuzzy Systems and Knowledge Discovery, pp. 1487–1492 (2015) 21. Qi, F., Yang, C., Liu, Z., et al.: Openhownet: an open sememe-based lexical knowledge base. arXiv preprint arXiv:1901.09957 (2019) 22. Li, S., Zhao, Z., Hu, R., et al.: Analogical reasoning on Chinese morphological and semantic relations. In: Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics, pp. 138–143 (2018)

Course Resources and Teaching Methods of Logistics Management Major Under Emerging Engineering Education Yong Gu, Ju Chen(B) , and Zhiping Liu School of Logistics Engineering, Wuhan University of Technology, Wuhan 430063, China

Abstract. The connotation of the emerging engineering education (3E) and its requirements to the logistics management major are analyzed. The existing course resources and teaching methods in colleges and universities are introduced. Based on the progressive training chain of student ability (introductory course - cognitive practice - basic courses of the major - professional practice - elective courses of the major- graduation design), according to the characteristics of introductory course, basic courses and elective courses, some suggestions on the application of course resources and teaching methods in theoretical teaching of logistics management major under 3E are put forward. According to the characteristics of 3E and logistics management major, the evaluation index system of theoretical teaching of logistics management major is established, and the evaluation process of theoretical teaching level is determined based on analytic hierarchy process. The aim is to make rational use of existing course resources and teaching methods, and promote the teaching of logistics management major. Keywords: 3E · Logistics management major · Course resources · Teaching methods

1 Introduction On February 18, 2017, the Ministry of Education held a seminar on the development strategy of higher Engineering education in Fudan University, and the participating universities reached a consensus on emerging engineering education (3E) in Fudan University. Subsequently, the Ministry of Education launched the project of “Research and Practice of 3E”. Logistics is not a direct major under 3E, but its latest development trend is closely related to it. The development of 3E will inevitably lead to a new round of reform and practice of logistics management major teaching [1].

2 3E and Its Requirements for Logistics Management Major 2.1 Connotation of 3E 3E is a major action plan to continuously deepen the reform of engineering education from the perspective of serving national strategy, meeting industrial demand and facing © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 434–443, 2021. https://doi.org/10.1007/978-3-030-80472-5_36

Course Resources and Teaching Methods of Logistics Management Major

435

future development [2, 3]. The connotation of 3E is to cultivate diversified, innovative and outstanding engineering talents in the future by following the guidance of moral education, taking coping with changes and shaping the future as the construction concept, and taking inheritance and innovation, crossover and integration, coordination and sharing as the main approaches [4]. From the perspective of discipline composition and subject connotation, pay attention to the intersections and integration between different majors, so that it can be competent to solve all kinds of complex engineering problems, economic and social problems. From the perspective of discipline professional development, build and develop 3E in a larger discipline space, professional space, problem space, social space and thinking space, so that it can deal with various future problems [2]. 2.2 Requirements of 3E on Logistics Management Major Logistics major belongs to the interdisciplinary comprehensive discipline, which fits in with 3E. Under 3E, the teaching and the cultivation of talents of logistics management major in colleges and universities need to solve the following problems: Integrating 3E into the teaching of logistics management major is an important issue, which has five characteristics: new concept of engineering education, new structure of disciplines and majors, new mode of talent training, new quality of education and teaching, and new system of classified development [5]. Compared with the traditional engineering talents, the future emerging industries and the new economy need highquality composite engineering talents with strong engineering practice ability, strong innovation ability and international competitiveness. This requires proactive changes in course system, educational and teaching methods [6]. How to make full use of existing course resources and teaching methods to improve the teaching effect of logistics management major. Logistics management talents need rich theory knowledge, practical ability, innovation ability and so on [7], this needs to adjust and change the traditional teaching mode, establishing the consciousness of the multi-disciplinary cross and fully apply theory to practice [8].

3 Current Situation of Course Resources and Teaching Methods 3.1 Course Resources The source of abundant teaching resources is the fundamental change of traditional teaching mode. The course resources of logistics management major in colleges and universities are mainly divided into online and offline aspects, as shown in Fig. 1. Online resources include video lessons, teaching software, electronic courseware, literature base, database, case base, test question database, etc. Video lessons are mainly watched through MOOC, Wisdom Tree, and other platforms. The teaching software of logistics management major includes ERP experiment system, international logistics experiment system, etc. Electronic courseware is mainly related to the content in class; The literature base mainly contains Chinese and foreign electronic books, journal articles and other literature resources, which can be retrieved through CNKI, school library and

436

Y. Gu et al.

other channels. Database can be obtained through library search, such as numerical database, economic statistics database and so on. Offline resources mainly include textbooks, reference books and experimental platforms established by colleges and universities. Textbooks and reference books are available from the school or college library. The relevant experimental platforms of logistics management major include crane experimental platform, virtual simulation teaching platform of logistics management, etc. video lessons teaching software electronic courseware Online resources

literature base database case base

Curriculum resources

test question database textbooks Offline resources

reference books experimental platform

Fig. 1. Course resources

3.2 Teaching Methods In order to achieve better results, flexible and diversified teaching methods should be adopted [9]. The teaching methods are mainly as follows. 1) General teaching method. It’s a traditional teaching method. In the teaching process, teachers usually demonstrate the graphics and animation with PPT, and explain the theoretical knowledge of their major in a way that is easy to understand. 2) Case teaching method. It’s usually more vivid and easier to motivate students than the general teaching method. The specific way is to integrate professional theoretical knowledge with intuitive cases, guide students to solve problems existing in the cases, and improve students’ ability to analyze and deal with problems. 3) Classroom practice method. Leave a small amount of time in class for students to study freely. The core knowledge of the course is further understood by the students, with fewer questions. Students submit their homework on the spot, and the teacher corrects it, so as to keep abreast of the students’ learning status and make targeted explanations. 4) Flipped classroom. Students use the teaching resources for self-study before class. Teachers will flip it in three ways according to the characteristics of knowledge

Course Resources and Teaching Methods of Logistics Management Major

437

points. Questions in class is applicable to the flipped classroom at the beginning of the course, so students can adapt to the self-learning process. By assigning exercises, hand in homework on the spot. This way can grasp the majority of students’ extracurricular learning, but more applicable to the theoretical basis of knowledge. Teacher gives the case, and students discuss the solution and explain it. It’s suitable for students to solve problems with comprehensive knowledge in the middle and later period of the course [10].

4 Suggestions on the Application of Course Resources and Teaching Methods Under 3E How to integrate 3E into the teaching of logistics management majors, and how to reasonably use course resources and teaching methods still need further research. Based on the progressive training chain for student ability (introductory course - cognitive practice - basic courses of the major - professional practice - elective courses of the major- graduation design), the application of course resources and teaching methods in the stage of theoretical teaching are analyzed (Fig. 2). The stage of theoretical teaching

Suggestions for application

Introductory course

Basic courses

Course resources: mainly in the form of pictures and videos, the teaching content is comprehensive Teaching methods :

combining theory with practice and teaching by many teachers Assessment method: test + course project

Connotation of 3E

Inheritance and Innovation

Elective courses

Course Course resources: depending on the course type, the content emphasis of similar courses is different Teaching methods: diversity Assessment methods: diversity

Crossover and integration

resources:

organize in-depth teaching content and pay attention to ability cultivation Teaching method: case teaching Assessment method: test + course project

Coordination and sharing

Diversity

Fig. 2. Integration of 3E and theoretical teaching

4.1 Introductory Course Introductory course is an enlightening course, and similar courses for logistics management also include modern logistics. Such courses are usually offered at the college entrance level and have fewer hours. This course is designed to enable students to establish a comprehensive understanding of the logistics discipline, so that students can have an intuitive understanding of the development history, knowledge structure, training objectives and requirements of the major as well as basic knowledge, typical technologies and specific applications related to logistics management at the initial stage of

438

Y. Gu et al.

university admission. The following are suggestions for the use of course resources and teaching methods at this stage. According to the course characteristics of introductory course, it is necessary to establish students’ comprehensive understanding of this subject in a way that is easy to understand and stimulate students’ enthusiasm. Therefore, teachers need to make a comprehensive introduction to the structure and system of the discipline in the form of pictures and videos from the perspective of practical life, without detailed and in-depth explanation of its composition, so that students can form a preliminary understanding. At the same time, through the online literature library and other channels, the new things and new technologies of the industry can be introduced to the students, so as to stimulate their interest in learning. Combine theory with practice and teaching by many teachers. Teaching by different teachers is helpful for students to understand the logistics discipline from multiple angles and aspects. The teaching process should be based on the combination of theory and practice, so students have a more intuitive knowledge and understanding of their major [11]. Adopt the assessment method of “test + course project”. The examination of students in the form of “test question + course project” is conducive to enhancing students’ mastery of basic knowledge. The arrangement of course project is conducive to guiding students to improve their ability to analyze, deal with and solve problems. 4.2 Basic Courses of the Major Basic courses include many courses, which can be roughly divided into two categories. The first category reflects the interdisciplinary integration of logistics management major, such as microeconomics, principles of management, operations research, etc., which are closely related to the logistics management major. The second category is specialized courses with the characteristics of the logistics industry, such as international logistics, logistics facilities and equipment, procurement management, etc., which aim to enable students to have a deeper understanding of the major. The course resources and teaching methods should be different according to the purpose and positioning of the two kinds of courses. The following are suggestions for the use of course resources and teaching methods at this stage. The course resources are different according to the course type, and the content emphasis of the same kind of courses is different. For the first category, it’s similar to most general subjects. Textbooks, reference books, and online course resources should be fully utilized. For the second category, in addition to the full use of basic course resources, the characteristic course resources of the major also should be used, such as the port crane platform and the virtual simulation teaching platform. Through visits and practices, students can form a more intuitive impression of the major. The content of similar courses is overlapping, so it’s necessary to coordinate the relevant knowledge points among similar courses. Take the “inventory control” as an example. In the course of “operation management”, the main emphasis is on the teaching of inventory control methods. In the course of “warehouse management”, the main emphasis is on the explanation of on-site operations and processes. In the course of “supply chain management”, the main emphasis is on the case analysis.

Course Resources and Teaching Methods of Logistics Management Major

439

Use diversified teaching methods. Interactive communication, combined with case teaching method, classroom practice, etc., to create a heuristic teaching atmosphere. When teachers explain a professional term or a phenomenon, they should specifically clarify the method of analyzing the problem and the steps to solve it. For example, in the “beer game” adopted in the supply chain management course, students play the roles of manufacturer, distributor, wholesaler, retailer and customer, which help students apply the knowledge to complex logistics problems [12]. Multi-disciplinary knowledge integration can improve the teaching effect of courses. Logistics management is an interdisciplinary subject. In order to deepen students’ understanding of it, students should also integrate the knowledge of other subjects, such as accounting, management, microeconomics and other knowledge. Diversified assessment methods should be adopted. Teachers can integrate the students’ mastery of knowledge into ordinary classes, assign small class assignments in stages. Let students make report on a topic, discuss a problem, and demonstrate a certain information technology in groups. 4.3 Elective Courses of the Major It refers to the professional improvement course that is closely related to or strongly related to the major. It’s usually set in the senior stage of undergraduate study, and its main purpose is in two aspects. According to the different interests of students, meet the personalized needs of students. Make the student to some professional knowledge understand more specialized, more refined, deeper, for graduation design, enter a higher school to read graduate school to lay the foundation. The following are suggestions for the use of course resources and teaching methods at this stage. Elective courses are usually offered at the advanced stage of the undergraduate course. Students already had systemic professional theory knowledge understanding. The depth of the course content should be increased. Make full use of video, animation, image and other multimedia forms. The abstract basic theory and unfamiliar technology will be presented to students more intuitively. Students can adopt the retrieval practice method, that is, students can recall the concepts they are interested in from the knowledge they have learned, which is more effective than simply reading textbooks [13]. The case teaching method is adopted. Cases are usually from the actual scientific research projects of the teachers. In the process of case analysis, students apply theories to practice by solving practical problems, and at the same time enable students to understand the system structure and main content of the course and grasp the main points of the course from a wide range of information. Adopt the assessment method of “test + course project”. As for elective courses, students have limited time, so it’s important to make full use of classroom teaching. The problem-oriented method is adopted to assign course project and papers to students at the end of the course, so as to improve students’ enthusiasm [14].

440

Y. Gu et al.

5 Evaluation of Theoretical Teaching 5.1 Construction of Evaluation Index System of Theoretical Teaching Theoretical teaching is the basis of talent training, and theoretical teaching should be advanced and forward-looking in combination with practice. Therefore, fuzzy evaluation, which can study data with uncertainty and fuzziness, is the most appropriate way to evaluate theoretical teaching. Analytic Hierarchy process (AHP) is a common method in the fuzzy evaluation model to determine the weight of each index layer. For each theoretical link of the training chain, it is necessary to carry out quantitative evaluation activities (Fig. 3). Organization of course content Course resources

Form of course resources Cultivation of students ability

Evaluation index system of theoretical teaching

Student participation Teaching methods

Degree of combination of theory and practice Teaching means Assessment of students ability

Assessment

Assessment of course content Form of assessment

Fig. 3. Evaluation index system of theoretical teaching

The elements in each layer are compared in pairs, and the comparison results are quantified according to the reference values listed in the Table 1. The first level is the evaluation index system of theoretical teaching (A). The second level is course resources (B1), teaching methods (B2) and assessment (B3). The third level is the organization of course content (C11), form of course resources (C12), cultivation of students’ ability (C13), student participation (C21), degree of combination of theory and practice (C22), teaching means (C23), assessment of students’ ability (C31), assessment of course content (C32) and form of assessment (C33). 5.2 Weight Calculation of Index System Introductory course, basic courses and elective courses of the major are in different stages of the training chain, and there are differences in emphasis and index weight of talents cultivation. Take the Introductory course as an example to calculate the weight.

Course Resources and Teaching Methods of Logistics Management Major

441

Table 1. Scale table Factor i compared to factor j

Quantitative values

Equally important

1

Slightly more important

3

More important

5

Highly important

7

Extremely important

9

The median of two adjacent judgments 2, 4, 6, 8

Through the expert scoring method, the importance degree of the second level index is compared in pairs. The judgment matrix of the second level is obtained. ⎤ ⎡ 135 (1) A = ⎣ 13 1 3 ⎦ 1 1 5 3 1 For matrix A, consistency test should be carried out to ensure that matrix A conforms to logic. After the consistency test, the factor weight of the fuzzy comprehensive evaluation is evaluated according to the triangular fuzzy judgment matrix. After processing the constructed fuzzy judgment matrix, the matrix B is obtained. The element calculation method of matrix B is shown in Formula 3. ⎡ ⎤ 0.652 0.692 0.556 B = ⎣ 0.218 0.231 0.333 ⎦ (2) 0.130 0.077 0.111 aij (3) bij = n  aij i=1

After the normalization operation of matrix B, the second-level normalized matrix W is obtained as follows. ⎡ ⎤ 0.633 W = ⎣ 0.261 ⎦ (4) 0.106 According to the above judgment matrix and its result value, the weight analysis of the second level can be seen, and similarly, the weight of the third level index can be obtained. According to the index weights of each level, the weights of the final index system are shown in Table 2. According to the weight of indicators at each level, the data distribution of evaluation indicators at each evaluation level can be counted with the help of teacher evaluation system, and students’ feedback to teachers and courses can be collected by designing

442

Y. Gu et al.

specific questions [15]. Through the analysis and calculation of the data, the comprehensive evaluation of the application effect of course resources, teaching methods and assessment methods in each stage can be obtained. Table 2. Weight table of theoretical teaching evaluation index system at all levels Evaluation index system of theoretical teaching

Course resources (B1) (0.633)

Organization of course content (C11) (0.431) Form of course resources (C12) (0.198) Cultivation of students’ ability (C13) (0.371)

Teaching methods (B2) (0.261)

Student participation (C21) (0.589) Degree of combination of theory and practice (C22) (0.215) Teaching means (C23) (0.201)

Assessment (B3) (0.106)

Assessment of students’ ability (C31) (0.410) Assessment of course content (C32) (0.357) Form of assessment (C33) (0.233)

6 Conclusion By analyzing the connotation of 3E and their requirements for logistics management major, it can be seen that 3E have the characteristics of inheritance and innovation, crossover and integration, coordination and sharing, etc., and how to integrate 3E into the teaching of logistics management major needs further study. The status quo of course resources and teaching methods in colleges and universities is analyzed. Teaching methods and course resources are abundant, but its concrete application still needs further study. Therefore, some suggestions on course resources and teaching methods research of logistics management major are put forward, from three aspects of introductory course, basic courses of major and elective courses of major, combining with 3E, and put forward the method of evaluating each stage of theoretical teaching. Acknowledgment. This research was financially supported by the Industry-University Cooperation Collaborative Education Program, the Ministry of Education of China through Grant No. 201802061010 and 201802061023 and also by Teaching Reform Research Program, Wuhan University of Technology through Grant No. w2017116.

Course Resources and Teaching Methods of Logistics Management Major

443

References 1. Han, T., He, J.: Research on the teaching of logistics and supply chain management in the context of emerging engineering education. Logist. Sci. Technol. 43(03), 175–177 (2020). (in Chinese) 2. Lin, J.: Emerging engineering education leading higher education reform. China High. Educ. (12), 40–43 (2017). (in Chinese) 3. Wang, X., Cao, Z.: Discussion reform of forestry panorama course teaching. Int. J. Mod. Educ. Comput. Sci. 12(29), 41–45 (2012) 4. Zhong, D.: Connotation and action of emerging engineering education. High. Eng. Educ. Res. (03), 1–6 (2017). (in Chinese) 5. Fatima, S., Abdullah, S.: Improving teaching methodology in system analysis and design using problem based learning for ABET. Int. J. Mod. Educ. Comput. Sci. 5(7), 60–68 (2013) 6. Sun, H., Cao, C., Song, Z.: Research on the cultivation of logistics engineering talents in the context of emerging engineering education. Logist. Eng. Manage. 41(4), 172–174 (2019). (in Chinese) 7. Fu, H.: Integration of logistics simulation technology and logistics learning factory in a twostage teaching method for logistics management courses. Int. J. Emerg. Technol. Learn. 12(9), 62–72 (2017) 8. Wei, X.: Discovery and practice of EDA experimental teaching reform. Int. J. Educ. Manage. Eng. 1(4), 41–45 (2011) 9. Chang, G., Zhang, Y., Yu, M., Yin, Z., Liu, M.: The application of teaching resources in the ideological and political construction of materials major course. J. High. Educ. (32), 181–184 (2020). (in Chinese) 10. Lu, L., He, Y., Li, X., Wang, L.: Teaching reform and practice of digital image processing course. Educ. Mod. 7(34), 64–67 (2020). (in Chinese) 11. Hadgraft, R.G., Kolmos, A.: Emerging learning environments in engineering education. Aust. J. Eng. Educ. 25(1), 3–16 (2020) 12. Lau, A.K.W.: Teaching supply chain management using a modified beer game: an action learning approach. Int. J. Logist. Res. Appl. 18(1), 62–81 (2015) 13. Anderson, L.C.: A survey of student engagement with multiple resources in an undergraduate physiology course: retrieve or look it up. Adv. Phys. Educ. 42(2), 348–353 (2018) 14. Alvarstein, V., Karen, J.: Problem-based learning approach in teaching lower level logistics and transportation. Int. J. Phys. Distrib. Logist. Manage. 31(7), 557–573 (2001) 15. Amjad, M., Linda, N.J.: A web based automated tool for course teacher evaluation system (TTE). Int. J. Educ. Manage. Eng. (IJEME) 10(2), 11–19 (2020)

Project-Oriented Course System Construction of Logistics Management Major Yue Liu, Yong Gu(B) , and Zhiping Liu School of Logistics Engineering, Wuhan University of Technology, Wuhan 430063, China

Abstract. As a comprehensive discipline with strong practicality, logistics management talents need to have a strong comprehensive ability. The problems in the cultivation of logistics management talents in colleges and universities are mainly caused by the unclear cultivation objectives and the unreasonable construction of the course system. The traditional course system of logistics management has been unable to meet the current needs of logistics management talent cultivation. By analyzing the current teaching situation of logistics management major, we summarize three reasons for the unreasonable construction of the course system, and propose a project-oriented course system of logistics management major with the goal of “solving complex logistics management problems”. In the aspect of improving the course system of logistics management specialty, some suggestions are put forward. Then from the practice courses and the theory courses two aspects set up the concrete courses. A course system evaluation method based on three stage DEA model is proposed. Keywords: Project-oriented · Course system · Ability of solving problems · Logistics management major

1 Introduction As an important part of macro-economy, logistics has attracted much attention from all sides [1]. The major of logistics management is a new major which adapts to the wide needs of the society. To cultivate compound talents with modern logistics management knowledge and skills is not only the need of modern logistics development, but also the basic talent cultivation goal of logistics management [2]. Under the background of education reform, the cultivation of logistics management talents needs to be integrated with new ideas, and the education reform also put forward new requirements for the cultivation of talents in colleges and universities [3]. At the same time, “Internet+ logistics” is changing the operation mode and efficiency of the traditional logistics industry from the aspects of technology, equipment and business model. Under this background, the logistics industry has an increasing demand for highquality complex logistics management talents [4]. In addition, under the background of high-quality economic development, the society has increasingly higher requirements on students’ comprehensive ability, which requires students to have strong basic professional ability, comprehensive professional ability, modern tool application ability and © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 444–453, 2021. https://doi.org/10.1007/978-3-030-80472-5_37

Project-Oriented Course System Construction

445

the ability of solving problems. The traditional logistics management course can’t meet this need well. Therefore, the course system of logistics management is in urgent need of improvement [5]. Based on the above background, we construct a project-oriented course system of logistics management with the goal of “solving complex logistics management problems”.

2 Current Situation of Logistics Management Major Teaching With the rapid development of economic globalization and Internet e-commerce, highquality logistics talents have become a scarce resource, and the cultivation of logistics management talents has also attracted wide attention from all sectors of society. At present, there are two main problems in the cultivation of logistics management talents. On the one hand, the cultivation objectives are not clear. On the other hand, the course system is not reasonable. Among them, the problems existing in course system construction can be subdivided into the following three points: 2.1 The Teaching Content of Specialized Courses Can’t Meet the Needs of Enterprises for Logistics Management Talents Driven by new technologies, the logistics industry needs to undergo structural adjustment and industrial upgrading, which will eventually lead to changes in logistics enterprise processes, organizational changes, the disappearance of traditional posts and the emergence of new ones. Post knowledge and skills change greatly, so that the old talent cultivation specifications no longer meet the requirements, the old talent cultivation program must reposition the direction of employment, re-determine the typical work tasks, re-determine the requirements of vocational ability [6]. In the face of the change of industrial structure, logistics management majors in colleges and universities are rarely able to make complete changes, but only make minor adjustments on the original basis, and the effect of talent cultivation is difficult to achieve the expected effect. 2.2 Students Have a Scattered Grasp of the Knowledge Points of the Course Because traditional teaching focuses on the introduction of theoretical definitions and technical applications, in the absence of project experience, students often don’t know how to apply specific knowledge comprehensively to solve practical problems. At the same time, students’ dependence on tools is too strong and their flexibility is poor. Although many students pass the course assessment, they don’t have a deep understanding of the mechanism behind the theory and framework, which leads to difficulties in subsequent self-study and in-depth research [7]. 2.3 The Teacher lays too much Stress on Theory and not Enough on Practice The course of logistics management has the problem of emphasizing theory and ignoring practice. At present, most domestic logistics management courses are mainly taught in

446

Y. Liu et al.

class. The emphasis on theory and the neglect of practice will result in the lack of practical ability and practical experience in students’ work, which will make them unable to adapt to the needs of their posts in time. Therefore, a large number of logistics management graduates don’t meet the social requirements for logistics talents [8]. To sum up, there is a lack of match between the cultivation of logistics management talents and the development of the logistics industry in China’s colleges and universities, and there is still a structural imbalance between the supply and demand of talents. Although some colleges and universities have made adjustments to the course system, they are still in the initial stage and can’t meet the demand for high-quality and innovative talents in the development of the logistics industry [9].

3 Project-Oriented Teaching Philosophy The “project-oriented” teaching method transforms the traditional teaching mode, which focuses on imparting knowledge, into the interactive teaching mode, which focuses on completing tasks and solving problems. The reoccurrence teaching is transformed into inquiry teaching to stimulate students’ active learning status, so that students can propose solutions and solve problems based on their own understanding of the current task and through the application of existing knowledge and experience. Project-oriented teaching method is more advantageous than traditional teaching method [10]. Project-oriented teaching method takes students as the main body, and students independently design the teaching content and implement the project plan. Through selflearning, self-practice and self-operation to acquire knowledge, students no longer passively accept knowledge, but actively explore and try. Project-oriented teaching method can effectively establish a new student-oriented teaching mode, and can formulate teaching tasks and carry out teaching project design according to the actual situation of students [11]. Project-oriented teaching method is not simply imparting knowledge, but under the careful organization and arrangement of teachers, students continuously improve their comprehensive ability and accomplishment through the process of accumulating theoretical knowledge and completing course experiments. This teaching mode emphasizes the process rather than the result. For students majoring in logistics management, a project-oriented teaching mode is set with the goal of cultivating students for four years to “solve complex logistics management problems”. The whole process of project target implementation is shown in Fig. 1. The course of logistics management involves many knowledge points and strong practicality. It is also widely applied in real life and production practice. Students are interested in the knowledge in this field and are easy to get involved in the learning process. The course design can be in accordance with the modern logistics enterprise post requirements and logistics management of the main knowledge points, the project content and work tasks detailed. The design of a teaching plan can be divided into several projects or themes according to the task process. Through the decomposition, optimization and reconstruction of knowledge, the course is closer to the reality [12].

Project-Oriented Course System Construction

447

Fig. 1. The implementation process of project-oriented teaching objectives

4 The Course System Construction of Project Oriented Logistics Management Major As a typical interdisciplinary subject, the major of logistics management has the following characteristics: 1) Students are required to have strong practical ability. Logistics management operation process has strong flexibility, lack of relatively unified implementation standards. Students majoring in logistics management often need to be able to arrange and coordinate flexibly according to the basic theoretical knowledge and practical skills learned in class. This requires that the teaching of logistics management should pay more attention to the cultivation of students’ practical ability. 2) Students are required to have strong system thinking ability. Generally speaking, logistics business involves many links. Therefore, students engaged in logistics management in the future should have certain systematic thinking and be able to comprehensively consider the influence of each link and make decisions. This requires teachers to be able to cultivate students in the classroom. According to the characteristics of logistics management, logistics management talents need to have strong comprehensive ability and problem-solving ability [13]. Therefore, we will build a project-oriented course system for logistics management with the goal of “solving complex logistics management problems”. 4.1 Suggestions on Improving the Course System of Logistics Management Major Based on the analysis of the current teaching situation of the logistics management major and the characteristics of the logistics management major, combined with the project-oriented teaching objectives, we put forward the following three suggestions:

448

Y. Liu et al.

4.1.1 Schools Should Define Their Teaching Orientation The course of logistics management should focus on the orientation of teaching and should combine the theoretical knowledge with the specific application and the latest technology according to the market demand, so as to impart the knowledge, cultivate the ability and cultivate the quality in line with the market demand. The comprehensive and systematic course system should be able to meet the individualized and differentiated learning needs of students and realize the docking of professional course construction and vocational ability on the basis of taking traditional teaching modules into account. The relationship between theoretical courses and theoretical courses, as well as the relationship between theoretical courses and practical courses, should be considered in the course setting to appropriately improve the proportion of practical courses [14].Course design can be selectively added to cultivate students’ problem-solving and teamwork skills. The proportion of course design assessment results should be appropriately increased and the proportion of pure theoretical knowledge assessment results should be reduced. 4.1.2 Schools Should Offer Interdisciplinary Courses Logistics management is a typical interdisciplinary integration of majors. Colleges and universities should take their own circumstances into consideration and offer interdisciplinary courses. For example, in addition to the core courses of logistics management major, mathematics courses such as operations research, probability theory and mathematical statistics can be offered. Engineering courses may include basic industrial engineering, introduction to Transportation Engineering, etc. Courses such as C language programming foundation and logistics information system can be offered in the computer class. For economic management, courses such as principles of management and microeconomics can be offered. The topic of graduation project can be targeted to set typical and multi-disciplinary complex logistics management problems, so that students can aim at a certain problem, through their own thinking, comprehensive use of knowledge in various fields and relevant tools, finally put forward solutions to the problem, cultivating students’ ability to solve complex logistics management problems. 4.1.3 Schools Should Implement Project-Oriented Teaching Methods Project teaching is the fundamental way to solve the problem that students have a relatively scattered grasp of the course knowledge, so that students can accumulate project experience in the process of project management and improve their comprehensive ability to analyze and solve problems. Project-oriented teaching methods can fully arouse students’ learning initiative and enthusiasm. Students use their existing knowledge and experience to acquire new knowledge and skills through practical learning. Both theory and practice should be taken into account in the assessment of students, so as to combine theory with practice. At the same time of mastering the professional theoretical knowledge can better complete the practical work given by teachers. At the same time, students not only master knowledge and skills in the process of completing tasks in cooperation with other students, but also cultivate and improve students’ teamwork ability.

Project-Oriented Course System Construction

449

4.2 Suggestions on the Course Setting of Logistics Management Major Looking ahead to the twenty-first century, with the continuous development of manufacturing industry and the progress of logistics technology, driven by intelligent technologies such as cloud computing, big data and the Internet of Things, the logistics industry is developing towards the direction of intelligent digitization [15]. Traditional logistics talents can no longer fully adapt to the rapidly changing requirements of logistics technology. Comparatively speaking, the focus of undergraduate teaching should be to cultivate the complex logistics senior management talents with theoretical knowledge and practical application of logistics management, so as to meet the requirements of senior logistics posts with higher requirements on comprehensive ability. Compound talents are strategic senior managers with strong comprehensive abilities.

Fig. 2. Theory course teaching hierarchy diagram

According to the above ability requirements, we provide courses for the major of logistics management major, which can be divided into theoretical courses and practical courses, as shown in Fig. 2 and Fig. 3. Through the study of theoretical courses, students majoring in logistics management can complete the process from preliminary understanding to reserve mastery of relevant theoretical knowledge. Through the study of practical courses, students can master the knowledge and skills of solving complex logistics management problems. In the hierarchy diagram of theoretical courses and practical courses, the course theory is gradually deepened from the inner layer to the outer layer, and the ability requirements of students are also gradually improved. Through the study of theoretical courses and practical courses in the course system, students majoring in logistics management can

450

Y. Liu et al.

Fig. 3. Practice course teaching hierarchy diagram

greatly improve their practical application ability, organization and management ability, innovation and creation ability, overall planning ability, project execution ability and logical reasoning ability, which is helpful for college graduates to quickly step into the society. 4.3 Evaluation of the Improved Course System Data Envelopment analysis (DEA) is a method of operations research and research on the boundary of economic production. This method is generally used to measure the productivity of some decision-making departments. In order to accurately evaluate the teaching effect of the improved course system, we use three-stage DEA model to eliminate the influence of environmental factors, random noise and management inefficiency, including the following three stages. The first stage: the effect evaluation of traditional DEA model. The input oriented Banker, Charnes and Cooper (BCC) model is selected to evaluate the teaching effect of the improved course system. Under the BCC model, the DEA model has the following results: 1) If θ = 1 and S + = 0, S − = 0, then the decision-making unit (DMU) is effective. 2) If the θ = 1, S + = 0 or S − = 0, then the DMU is considered as weak effective. 3) If the θ < 1, then the decision unit is invalid. Wherein, technical efficiency (TE) of BCC model = pure technical efficiency (PTE) × scale efficiency (SE). The second stage: construct the similar SFA model. The relaxation variable of the input variable in the first stage is decomposed into a function containing three independent variables, namely environmental factor, random noise and management inefficiency

Project-Oriented Course System Construction

factor, and its expression is constructed by formula (1):   Smi = f zi ; β m + vmi + umi

451

(1)

Where, m represents the input variable; i represents the decision unit; Smi represents the relaxation variable input in item m of the ith DMU; f(zi ; β m ) denotes the influence of environmental factors on input relaxation variables, usually f(zi ; β m ) = zi ∗ β m , zi is the observed environmental factor, β m is the corresponding coefficient of environmental factor to be estimated. vmi + umi is the mixed error term. vmi represents the influence of random interference on input relaxation variable. umi reflects the impact of management inefficiencies on input relaxation variables. Then the parameters such as β m , σ 2 , and γ were estimated according to the SFA estimates, and then vmi and umi were calσ2

um culated according to the above parameters. Defining γ = σ 2 +σ 2 , management factors um vm dominated when γ → 1.When γ → 0, the random factor is dominant. Based on formula (1), the new input variable at the same level is calculated by formula (2):     xmi = xmi + max zi β m − zi β m + [max{vmi } − vmi ] (2) 



Where, xmi and xmi respectively decision-making unit of input values before and after the adjustment, i indicates the decision making units, m indicates redundant input variables; max{zi β m } − zi β m adjusts the influence of environmental factors. max{zi β m } indicates that under the worst environmental conditions, other decision making units are adjusted according to the standard. More input is added in good conditions, and less input is added in bad conditions, so all decision making units are adjusted to the same environmental level. max{vmi } − vmi is to adjust the random noise. The principle is the same as above, that is, all decision making units have the same randomness. The third stage: adjusted input values xmi as input, once again, the traditional DEA model is used to calculate the relative efficiency of decision making units, the more can objectively reflect the efficiency value of the improved teaching effect of course system [16]. Input-output indexes are set as follows: Grade I consists of input indexes and output indexes. The input indexes include course content, teaching staff, instructional resources and courses. The output indexes include practical application ability, organization and management ability, innovation and creation ability, overall planning ability, project execution ability and logical thinking ability. The entropy weight method can be used to determine the weight value of each index. 

5 Conclusion The project-oriented concept is a typical student-centered teaching concept, which can effectively enhance students’ learning motivation and better cultivate their thinking and problem-solving ability. With the continuous development of modern society, the influence of the logistics industry is gradually expanding, and the requirements for logistics management talents are also gradually improving. Based on logistics management as an example, we

452

Y. Liu et al.

aimed at the problems of logistics management professional talent cultivation, in order to “solve the problem of complex logistics management” as the goal put forward building the project oriented course system and course of specific recommendations, course cultivation system can be improved better cultivate the students’ comprehensive ability. Acknowledgment. This research was financially supported by the Industry-University Cooperation Collaborative Education Program, the Ministry of Education of China through Grant No. 201802061010 and 201802061023 and also by Teaching Reform Research Program, Wuhan University of Technology through Grant No. w2017116.

References 1. Gan, W., Zhang, T., Zhu, Y.: On RFID application in the information system of rail logistics center. Int. J. Educ. Manag. Eng. 3(2), 52–58 (2013) 2. Dong, Y.: Modular teaching mode design of logistics management based on comprehensive ability improvement. Logist. Eng. Manag. 41(12), 180–183 (2019). (in Chinese) 3. Yu, L.: Research on the “Cooperative Education” model cultivating in higher vocational education. Int. J. Educ. Manag. Eng. 2(1), 521–524 (2012) 4. Zhang, Y., Zhang, S., Shao, Q.: Research on the optimization of curriculum system of logistics engineering major under internet + background. In: Proceedings of the 2018 4th Annual International Conference on Modern Education and Social Science (MESS 2018) (2018) 5. Fatima, S., Abdullah, S.: Improving teaching methodology in system analysis and design using problem based learning for ABET. Int. J. Mod. Educ. Comput. Sci. 5(7), 60–68 (2013) 6. Xia, W., Xia, Q., Liu, W.: Research on the collaborative training mechanism and curriculum system reform of logistics management talents. In: 2019 3rd International Conference on Education, Economics and Management Research (ICEEMR 2019) (2019) 7. Xie, C., Zhao, L., Zeng, L.: Practice exploration of project-based teaching mode of logistics management centered on vocational ability cultivation - a case study of e-commerce major in independent colleges. Econ. (8), 175–176 (2020). (in Chinese) 8. Fan, Z., Zhao, Y., Hu, Y.: Construction and practice of logistics management course based on OBE concept. China Storage Transp. (02), 114–115 (2020). (in Chinese) 9. Chen, J.: Practical exploration of college logistics management course teaching from the perspective of deep integration of industry and education. J. Jilin Normal Univ. Eng. Technol. 35(07), 75–77 (2019). (in Chinese) 10. Wang, X.: Application of “project-oriented and task-driven” teaching method in warehouse management. Educ. Teach. Forum (15), 144–145 (2015). (in Chinese) 11. Qu, R.: Design and implementation of project-oriented and task-driven teaching model—take logistics strategy consulting as an example. Lifelong Educ. 9(6), 243 (2020). https://doi.org/ 10.18282/le.v9i6.1351 12. Bian, H.: Course reform and practice of “logistics information technology” based on PBL. Educ. Careers (28), 98–101 (2015). (in Chinese) 13. Meng, L., Zheng, N., Wang, Q.: The exploration of the curriculum reform of warehousing and distribution management based on the demand of application-oriented logistics management talents. In: International Conference on Modern Educational Technology and Innovation and Entrepreneurship (ICMETIE 2020) (2020) 14. Pan, Y., Liu, D., Li, L.: Research and practice of practice course system construction of logistics management major in applied technology colleges and universities. In: Proceedings of

Project-Oriented Course System Construction

453

2018 4th International Conference on Education & Cultivation, Management and Humanities Science (ETMHS 2018). Institute of Management Science and Industrial Engineering: Computer Science and Electronic Technology International Society, vol. 3 (2018) 15. Yang, M., Mahmood, M., Zhou, X., Shafaq, S., Zahid, L.: Design and implementation of cloud platform for intelligent logistics in the trend of intellectualization. China Commun. 14(10), 180–191 (2017) 16. Yang, R., Huang, S., Wang, Y.: Evaluation of knowledge exchange efficiency in online health communities based on three-stage DEA model. Inf. Theory Pract. 43(10), 122–129

Analysis on the Venations, Hotspots and Trend of China’s Education Informatization Research in the Post-COVID-19 Era Xiaofen Zhou1 , Yi Zhang2(B) , and Yanan Wang2,3 1 School of Logistics, Wuhan Technology and Business University, Wuhan 430065, China 2 School of Business, Sichuan University, Chengdu 610064, China 3 School of Economics and Management, Tibet University, Tibet 850000, China

Abstract. The outbreak of COVID-19 has accelerated the development of educational informatization in China. This paper uses the CiteSpace to analyzes the statistical results of the educational information. The results show that: from the perspective of research venations, the research content, perspective and topic selection of educational informatization are gradually deepening; from the perspective of research hotspots, artificial intelligence and educational modernization, educational equity and digital divide, educational governance and educational quality evaluation system, information-based teaching and higher education are the hot issues in the field of educational informatization research; The frontier trends of research are manifested in five aspects: research on the coordination of education poverty alleviation and regional education development, research on the coupling of online teaching and information literacy, discussion on the relationship between education governance system and governance capabilities, digital education and the optimal supply of educational resources, research on the relationship between artificial intelligence and smart education. Research conclusions can provide reference for subsequent related research. Keywords: Educational informatization · COVID-19 · Research frontier · Research hotspot

1 Instruction Education informatization has the unique advantages of breaking through the spacetime constraints, rapid replication and dissemination, and rich presentation means. It is an effective means to promote education equity and improve the quality of education [1], educational informatization is also the basic connotation and significant feature of educational modernization [2]. The sudden outbreak of the COVID-19 in early 2020 has had a great impact on China’s educational informatization. During the anti -epidemic period, for the life and health safety of teachers and students, the Ministry of Education issued the “notice on supporting education and teaching with information during the epidemic prevention and control period” in February 2020, which put forward the overall requirements for the teaching form during the epidemic period [3]. This reflects the important value of © The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 454–465, 2021. https://doi.org/10.1007/978-3-030-80472-5_38

Analysis on the Venations, Hotspots and Trend

455

information technology in the field of education, indicating that China has made remarkable achievements in the field of education informatization, but also exposed a series of problems in the process of teaching [4]. Some literatures show that the single practice mode, the lack of online course resources, the lack of information literacy of participants and the weak infrastructure are the key factors restricting the teaching effect, the resulting “digital divide” has also led scholars to think about the issue of educational equity [5, 6]. The occurrence of the epidemic situation makes the information-based teaching appear in the form of emergency. However, in the post epidemic era, how to develop educational informatization, how to realize resource sharing and educational equity, and how to realize educational modernization through educational informatization need to be considered [7]. Under the strategic background of realizing national education modernization with educational informatization, domestic scholars have made in-depth interpretation of educational informatization from multiple perspectives and levels [8]. The development of any subject is closely related to the domestic and international environment, so is the development of educational informatization. In the post epidemic era, the normal development of education has been broken [9]. Some scholars have proposed that the emergence of the COVID-19 will be a turning point in the development of education informatization [10]. However, the development of educational informatization presents new trends and characteristics, and how to overcome the current challenges, there is still a lack of necessary literature guidance. Therefore, the author attempts to use CiteSpace to systematically and comprehensively sort out the path and venations of China’s education informatization research, explore and analyze the new hot spots and new trends of educational informatization research since the outbreak of the epidemic, so as to offer some suggestions for future research.

2 Time Distribution of Educational Informatization Research in China 2.1 Data Collection The literature samples used in this study were retrieved from CNKI data platform. According to Bradford’s law, the core regional literature is more valuable and authoritative. Therefore, in order to ensure the scientificity and representativeness of the literature, this paper selects “advanced retrieval” in CNKI and sets “CSSCI” as the constraint condition. The setting conditions of literature retrieval and the details of literature elimination are shown in Table 1. 2.2 Time Distribution of Educational Informatization Research in China The number of published literatures is an important indicator to measure the development trend of a research field in the given time period, which is of great significance to analyze the development trend and forecast the future trend of this field. In China, the earliest research literature about education informatization can be retrieved is in the late 1990s. In order to make a more in-depth and comprehensive analysis of the evolution of

456

X. Zhou et al. Table 1. Conditions of document retrieval and literature elimination

Search options

Condition setting

Types of literature elimination

Elimination quantity

Column selection

Journal

Conference information 302 articles and summary

Subject

Education Informatization

Policy interpretation

37 articles

Dates

From 2000 to 2020 Special topic interview

30 articles

Search form

Fuzzy

Repetitive and other invalid literatures

303 articles

Classification

CSSCI

The effective rate of literature: 80.99% (=2863/3535)

Note: the retrieval time is Oct. 4, 2020

educational informatization research in China since the beginning of the new century, this paper makes statistics on the publication time of 2863 effective literatures, and the statistical results are shown in Fig. 1. In the past 20 years from 2000 to 2019, the number of educational information published in China showed a steady growth trend, and in 2016, the peak of the number of publications was issued (the peak value was 217). Generally speaking, the process of educational informatization research in China can be roughly divided into three stages. The first stage (before 2003): this stage is the initial stage of educational informatization research in China. In the 1990s, the Ministry of education and the State Council issued documents successively, emphasizing the important role of educational technology and educational informatization in quality education. Nan (2002), a leading academic scholar, pointed out in a seminar that educational informatization is a process of promoting educational modernization by using modern information technology to continuously improve education and teaching, cultivate and improve students’ information literacy. As a result, the theory and research of educational informatization also began to show a rising trend. The second stage (2003–2016): this stage is the rapid growth period of educational information research in China. Starting from the establishment of the Education Information Office of the Ministry of education in 2006, the theoretical research on educational informatization in China has also appeared a rapid growth period. The third stage (2017–): This stage is a new period for the development of China’s education informatization from 1.0 to 2.0. With the release of “education informatization 2.0 behavior plan” in 2018 as a sign, the research results of education informatization in China are increasingly abundant. Among them, there are not only the summary of the development process and successful experience of education informatization [11], but also research on the formulation of education informatization standards [12]. In terms of the number of literature, since 2000, the relevant research results showed an upward trend, and after reaching the peak in 2016, it showed a downward trend, and

Analysis on the Venations, Hotspots and Trend

Numbers of papers published article

250

217 201 210 186193195 177 154 154 139148 135 119121115

200 150 100 50 25

86 49 54

101

457

84

0

Time of publication

year

Fig. 1. Statistics of the number of papers published in education informatization from 2000 to 2020

the research enthusiasm decreased slightly; from the perspective of literature quality, the research depth and difficulty in the field of educational informatization are increasing. However, under the influence of the COVID-19, the retrieval data from 2019 to October 2020 showed an upward trend. Therefore, it is of great significance to systematically analyze the current research venations and hot issues of education informatization, combined with the impact of the epidemic on education informatization, and provide guidance for education informatization in the Post-COVID-19 Era.

3 Analysis of the Venation, Hot Issues and Frontiers of Educational Informatization in China 3.1 Analysis of the Venation of Research Bibliometrics is a quantitative analysis method based on the theoretical basis of mathematics and statistics, taking the external characteristics of scientific and technological literature as the object [13, 18, 19], which helps us to understand the development status and laws of various disciplines. CiteSpace is a widely used bibliometric research tool. This paper uses CiteSpace as research tools to analyze the venations, hot spots and trends of education informatization research. With the help of CiteSpace analysis tool, the evolution venations map with the educational informatization is drawn, as shown in Fig. 2. Notes: The original version of Fig. 2 is Chinese. At the request of the editorial ICAILE2021, the Fig. 2 was translated into English. If you need the original picture, please contact the Corresponding Author. Combining the stage division of the second part, as well as the research venations of education informatization (Fig. 2) and the keyword co-occurrence table from 2000 to 2020 (Table 2), the research venations of education informatization is sorted out.

458

X. Zhou et al.

Fig. 2. The research venations of educational information

The first stage is the initial stage (before 2003): the key words of this stage include information technology (177), educational technology (144), information education (65), primary and secondary school teachers (26), information technology resources (23), etc. It shows that the research in this stage mainly focuses on the theoretical exploration and basic application of educational informatization. The second stage is the rapid growth period (2003–2016). The key words of this stage are information literacy (57), basic education (87), education equity (42), higher education (50), wisdom education (83), etc. This stage is the explosive growth stage of educational informatization research. It shows that China’s investment in infrastructure has made some achievements in the early stage. The education informatization has been widely concerned by scholars, and its content system has also been enriched and improved. Scholars’ focus has shifted to flexible fields, such as student literacy, education equity, etc. The third stage is the transition period from 1.0 to 2.0 (2017). The key words of this stage include educational informatization 2.0 (42), artificial intelligence (31), new era (28), Internet plus education (20), supply side reform (19), etc. This stage is a new period for the transformation and upgrading of educational informatization in China. In this period, the products of a new round of scientific and technological revolution represented by artificial intelligence (AI) and 5G were applied in the field of education. How to enter the era of 2.0, realize the supply side reform of education with educational informatization and realize the modernization of education has become the main problem in this period. As a result, smart education, digital education, education reform, precision teaching, education equity and other fields have become research hotspots and new trends in the field of education informatization.

Analysis on the Venations, Hotspots and Trend

459

Table 2. Keyword co-occurrence table from 2000 to 2020 (condition: Centrality > 0.01 and rank ≥ 5) Stage

Sort

Centrality

Count

Keyword

First Stage (2000–2002)

1

0.21

177

Informationization

2

0.20

144

Educational technology

3

0.12

65

Information education

4

0.07

26

Primary and secondary school teachers

5

0.05

23

Information technology resources

1

0.15

57

Information literacy

2

0.12

87

Elementary education

3

0.09

42

Education equity

4

0.08

50

Higher education

5

0.05

83

Wisdom education

1

0.02

42

Education informatization 2.0

2

0.02

31

Artificial intelligence

3

0.02

28

New era

4

0.01

20

Internet plus education

5

0.01

19

Supply-side reform

Second Stage (2003–2016)

Third Stage (2017–)

3.2 Analysis of the Venation of Research Key words are a high generalization of the theme of the article, the essence and core of an article, and can reflect the research hotspots in a period of time from a macro perspective [14]. Figure 3 shows the results of Citespace tool’s analysis of relevant literature keywords. Combining the analysis of the keyword map in Fig. 3 and the research of related literature, the author believes that the hot spots of China’s education informatization research mainly include four aspects: artificial intelligence and modern education, education equity and digital divide, education governance and education quality evaluation system, informatization teaching and higher education. Notes: The original Version of Fig. 3 is Chinese. At the request of the editorial ICAILE2021, the Fig. 3 was translated into English. If you need the original picture, please contact the Corresponding Author. Notes: #0 - artificial intelligence; #1 - education informatization; #2 - targeted poverty alleviation by education; #3 - performance evaluation; #4 - information literacy; #5 information education;#6 - information leadership; #7 - Educational equity; #8 - higher education informationization; #9 - information teaching, #10 - innovative application; #11 - model.

460

X. Zhou et al.

Fig. 3. The map of keywords for China’s education informatization research

It can be found from Fig. 3 that the research hotspot of artificial intelligence and education modernization includes two clusters: #0 artificial intelligence and #1 education informatization. In the cluster of artificial intelligence, the key words with high centrality are intelligent education, cloud computing, education informatization 2.0, big data, etc. The key words included in the cluster of educational informatization are teaching method innovation, talent training mode, system framework, etc. Artificial intelligence reshapes the characteristics and style of the teaching environment, realizes the personalization of the nature of learning, the flexibility of learning time, the diversification of teaching teachers, and the hybridization of learning methods, which provides new possibilities for the reform and development of education [15]. Online teaching realizes the combination of talent training, intelligent evaluation system and individual needs, which provides a platform for the realization of personalized education. Under the concept of artificial intelligence driving educational reform, it is the requirement of educational reform in the intelligent era to meet the learning requirements of learners at any time, any place and in any way. In general, the reshaping effect of artificial intelligence on modern education has become a current research hotspot. The research hotspot of education equity and digital divide includes three clusters of #2 targeted poverty alleviation by education, #4 information literacy and #7 education equity. From Fig. 3, it can be found that the keywords for #2 targeted poverty alleviation by education include balanced development of education, co-construction and sharing, high-quality educational resources, etc.; #4 information literacy includes digital educational resources, deep integration, intelligent technology, etc.; #7 Educational equity includes information technology, teacher professional development, education balance, etc. Promoting educational equity with information resources is an important measure

Analysis on the Venations, Hotspots and Trend

461

to realize educational modernization in China. But at the same time, the implementation and development of education equity encounters the “first digital divide” due to the different levels of regional economic development caused by the different holdings of information technology equipment. And the “second digital divide” caused by the use of information technology and the degree of skill difference. Give full play to the important role of education informatization in education equity, and solve the digital divide in the use process has become a research focus. The research hotspot of education governance and education quality evaluation system mainly includes three clusters: #3 performance evaluation, #6 information leadership and #11 model. Among them, the key words of #3 performance evaluation include index system, evaluation, teaching methods, professional development, etc.; the key words of #6 information leadership include influencing factors, core literacy, quality literacy and so on; the key words of #11 model include educational technology, teacher training, talent training, intelligent learning, etc. In the environment where the new round of intelligent technology is widely used in teaching, China’s education concept and evaluation of education quality have become a hot topic, which is an inevitable process to realize the modernization of education governance and an important way to avoid the inconsistency between teaching methods and evaluation system. The research hotspot of information teaching and higher education mainly includes four clusters: #5 information education, #8 higher education informationization, #9 information teaching, #10 innovative application four clusters. The keywords of #5 information education include distance education, information education, resource construction, etc.; #8 Higher education informationization includes education cloud, innovative practice, teaching informationization, etc.; #9 Information education includes informationization construction, teaching innovation, regional type, etc.; #10 innovative applications include MOOC, education publishing, digital education, etc. Higher education is the key part of national training of senior talents. The integration of information technology and higher education is the key to realize the modernization of education and an important part of the development of education informatization. 3.3 Analysis of Research Frontier Problems in Post-COVID-19 Era The analysis of research frontiers can explore the focus topics and future research trends in the field of educational informatization. Some scholars pointed out that the COVID-19 epidemic is a dividing line, and world education will be divided into “pre- COVID-19 era” and “post- COVID-19 era” [10]. The outbreak of the epidemic is a major test of the development of education informatization in China. It is an unprecedented large-scale socialization experiment in which information technology and education and teaching are deeply integrated. It also has a huge impact on the development of education informatization in my country [16]. This section uses the Burstness function in Citespace to draw the burst table (shown in Table 3) with the burst time in 2018–2020 based on the intensity of burst terms with 2019 as the intermediate time to determine the latest evolution trend in the field of educational informatization. From Table 3, it can be seen that the burst terms of education informatization 2.0, artificial intelligence, smart campus, intelligent education and teaching reform were in the forefront before the outbreak of COVID-19 epidemic. after the epidemic, emerging

462

X. Zhou et al.

Table 3. Ranking of burst terms in China’s education informatization from 2018 to 2020 (top 10 of burst intensity) Burst terms

Intensity

Burst time

Burst terms

Intensity

Burst time

Educational Information 2.0

21.9959

2018–2020

Education equity

4.2958

2019–2020

Artificial intelligence

14.0759

2018–2020

Online 3.4073 learning space

2019–2020

Smart campus

4.3165

2018–2020

Educational governance

3.0326

2019–2020

Intelligent education

4.1435

2018–2020

Precision teaching

2.5132

2019–2020

Teaching Reform

3.3939

2018–2020

Smart classroom

2.2531

2019–2020

New Era

3.1562

2018–2020

Internet+

2.2396

2019–2020

Supply-side reform

2.5241

2018–2020

Digital education

1.8843

2019–2020

Digital Education Resources

2.2943

2018–2020

Personalized learning

1.8843

2019–2020

Educational Information Service

1.2612

2018–2020

Intelligent Technology

1.7540

2019–2020

Teaching innovation

1.1360

2018–2020

University Teachers

1.6235

2019–2020

words such as education fairness, online learning, education governance, and precision education were in the forefront, which shows that the epidemic has a greater impact on the development of informatization education. With the impact of the epidemic, the shortcomings of education informatization in China have emerged. For example, the single practice mode, the lack of online course resources, the lack of information literacy of participants and the weak infrastructure have become the obstacles restricting online education. As a result, the focus of attention of scholars began to focus on the fields of education equity, online learning space, education governance, and precision teaching. Based on the burst terms in Table 3 and combined with key literature, this paper summarizes the research frontiers of education informatization in the post epidemic era. First, research on the coordination mechanism of education poverty alleviation and regional education development. Second, research on the coupling of online teaching and information literacy. The specific design includes information literacy evaluation index construction, teacher information literacy, and student information literacy cultivation. Third, research on education governance system and governance capabilities. Including national governance system, regional teaching governance system, school education governance system research. Fourth, optimizing the supply of digital education and educational resources. Including development concepts, design ideas, evaluation systems, coupling and coordination relationships, etc. Fifth, artificial intelligence and

Analysis on the Venations, Hotspots and Trend

463

wisdom education research. Including 5G, blockchain, interactive relationships, policy incentives, implementation paths, etc.

4 Discussion The research on the development process of education informatization should not only start from the actual situation of China’s education development, but also combine with the use of modern information technology in education and teaching [18]. Generally speaking, since the concept of education informatization was proposed in China at the end of last century, domestic scholars have conducted in-depth research on education informatization. In the past 20 years, the content of the research on educational informatization in China has changed from theoretical research to practical application, from case study to regional promotion, from information teaching to digital teaching, from single subject research to multi-disciplinary comprehensive research; The topic of the research has shifted from the “material-based” education informatization construction and education consumption to the information literacy and “people-oriented” education equity and “digital divide”[19]; from the perspective of horizontal research, the shift from focusing on the construction of information resources in developed areas to the construction of information resources in impoverished areas and ethnic areas; from the perspective of vertical research, the shift from national-level institutional and framework research to micro-level regional education informatization. 1) This paper systematically analyzes the research status in the field of educational informatization from three aspects: basic situation, evolution venations and hotspots. At the same time, this paper analyzes the new hotspots and new trends that domestic scholars have paid attention to in the field of education informatization since the outbreak of the COVID-19 epidemic [20, 21]. At the same time, this paper analyzes the new hot spots and new trends of domestic scholars in the field of education informatization since the outbreak of the COVID-19 epidemic, which provides a reference for the selection of subsequent research directions. 2) However, it is worth noting that, subject to CiteSpace tools, this paper does not conduct key literature and collaborator analysis. Whether the focus of the important authors is consistent or not, and the characteristics of the cooperation network of different scholars need to be further studied. In addition, The definition of the concept of “post COVID-19 era” is not clear. This study only selects the burst terms of 2019–2020 to analyze the frontier issues of the research, the frontier conclusion of education informatization research is only for reference.

5 Conclusion This study uses 2863 CSSCI-level literatures on education informatization retrieved from CNKI from 2000 to 2020 as effective samples, and analyzes the statistical results of the research venation, keyword co-occurrence, keyword map and emergent words by CiteSpace tool [17]. The basic situation, evolutionary venations, research hotspots and research frontiers of the research field of education informatization in my country have been systematically combed, and the following conclusions have been obtained:

464

X. Zhou et al.

1. From the perspective of time distribution, China’s education informatization research originated at the end of the 1990s. After entering the new century, combined with the statistics of the number of papers published, it can be divided into the initial stage (before 2003), the rapid growth period (2003–2016), and the transformation and development period (education informatization 1.0 to 2.0). 2. From the perspective of research venations, the research of educational informatization has different concerns in different stages. On the whole, the research content, perspective, level, topic selection and other aspects show the characteristics of gradual deepening. 3. From the perspective of research hot issues, artificial intelligence and education modernization, education equity and digital divide, education governance and education quality evaluation system, informatization teaching and higher education are the hot issues in the field of education informatization research. 4. From the perspective of the frontier trend of research, affected by the COVID-19 epidemic, research in the post COVID-19 era focuses more on five aspects: research on the coordination mechanism of education poverty alleviation and regional education development, research on the coupling of online teaching and information literacy, research on education governance system and governance capabilities, research on the relationship between capabilities, digital education and the optimal supply of educational resources, and research on the relationship between artificial intelligence and smart education.

Acknowledgment. This project is supported by Educational Science Planning in Hubei Province in 2018(Research on Online Course Teaching Design of Applied Universities, 2018GB121); Educational Reform Project of Tibet Autonomous Region (JG019-31); the Teaching Team Project of Hubei Province (Teaching team of the construction of smart logistics curriculum system).

References 1. Zhang, C.: The construction limitation and practice breakthrough of the wisdom class in the education informationization. Soc. Sci. Nanjing (4), 136–141+156 (2020). (in Chinese) 2. Xu, Y.: A new education mode of engineering specialty graduate students—co-founding by college and enterprises, uniting of production, learning and researching. Int. J. Educ. Manag. Eng. 2(12), 590–593 (2012) 3. Wei, X.: Discovery and practice of EDA experimental teaching reform. Int. J. Educ. Manag. Eng. 1(4), 41–45 (2011) 4. Liu, J.: Research on strategies and patterns of research teaching. Int. J. Educ. Manag. Eng. 2(12), 67–72 (2012) 5. Lei, W., Huang, X.: Problems in the development of online education under the COVID-19 epidemic. Educ. Res. Exp. 41(8), 109–114 (2020). (in Chinese) 6. Dwivedi, Y.K., Hughes, D.L., Coombs, C., et al.: Impact of COVID-19 pandemic on information management research and practice: transforming education, work and life. Int. J. Inf. Manag. 1–20 (2020) 7. Yu, L.: Research on the “Cooperative Education” model cultivating in higher vocational education. Int. J. Educ. Manag. Eng. 2(1), 521–524 (2012)

Analysis on the Venations, Hotspots and Trend

465

8. Zhang, L., Wei, D., Huang, L.: The research on the program of educational technology ability building for vocational education teachers. Int. J. Educ. Manag. Eng. 2(11), 46–51 (2012) 9. Feng, J.: Reconstructing the new normal of education in the post-epidemic period. China Educ. Technol. (5), 1–6 (2020). (in Chinese) 10. Li, Z.: The post-epidemic era of basic education is a new era of “Online and Offline Blended Teaching Mode”. J. Chin. Soc. Educ. (5), 5 (2020). (in Chinese) 11. Cheng, L.: The analysis and research synthesis of the current situation of balanced allocation of educational information resources in China. Inf. Sci. 35(7), 170–176 (2017). (in Chinese) 12. Wu, D., Wang, Y., Peng, X.: Review on standards research of ICT in education. E-education Res. 40(1), 45–51+76 (2019). (in Chinese) 13. Zhu, L., Meng, X.: The comparative study on bibliometric method and content analysis method. Libr. Work Study (6), 64–66 (2013). (in Chinese) 14. Wei, F., Grubesic, T.H., Bishop, B.W.: Exploring the GIS knowledge domain using citespace. Prof. Geogr. 67(3), 374–384 (2015) 15. Zhu, Y., Xu, Z., Lubai, et al.: Written talk on “Artificial Intelligence and Future Education (Part I). J. East China Normal Univ. (Educ. Sci. Ed.) 35(4), 80–87 (2017). (in Chinese) 16. Zhou, H.: Towards the new stage of education informationization development in the new Era. J. Chin. Soc. Educ. (10), 5 (2020). (in Chinese) 17. Chen, C.: CiteSpace II: detecting and visualizing emerging trends and transient patterns in scientific literature. J. Am. Soc. Inform. Sci. Technol. 57(3), 359–377 (2006) 18. Yang, L.: The valid connection between the information service of the university library and the education informationization. Econ. Res. Guid. 33, 258–259 (2009). (in Chinese) 19. Zhao, S.: Parental education and children’s online health information seeking: beyond the digital divide debate. Soc. Sci. Med. 69(10), 1501–1505 (2009) 20. Morgan-Daniel, J., Ansell, M., Adkins, L.E.: COVID-19 patient education and consumer health information resources and services. J. Consum. Health Internet 24(3), 302–313 (2020) 21. Wang, Y.C.: Information behavior of parents during COVID-19 in relation to their young school-age children’s education. Ser. Librarian 1–16 (2020)

Author Index

A Aitchanov, Bekmurza, 162 Aitchanov, Tley, 162 Akatayev, Nurbol, 147, 326 Akhmetov, Berik, 326 Aleksander, Marek, 147 Alifov, Alishir A., 187 Avkurova, Zhadyra, 117, 147 B Baimuratov, Olimzhon, 162 Balykhina, Hanna, 89 Bauyrzhan, Madina, 117 Bohaienko, Vsevolod, 15 Bomba, Andrii, 75 Brych, Vasyl, 138 Buriachok, V., 374 C Chaika-Petehyrych, Liliia, 363 Chen, Ju, 434 D Demchuk, Olena, 89 Dobrotvor, Ihor, 304 Dong, Pan, 411 Dorovskaja, Irina, 102 Dorozhynskyy, S., 127 F Fedushko, Solomiia, 56, 256 Fesenko, A., 127 Fomishyna, Vira, 363

G Gagnidze, Avtandil, 117 Gavrilenko, Olena, 232 Gizun, Andrii, 147 Gnatyuk, Sergiy, 117, 326 Gobov, Denys, 208 Gu, Yong, 434, 444 Guo, Lei, 175 Gusev, Victor, 102 H Honcharenko, Oleksandr, 315 Hriha, Vladyslav, 147 Hu, Z. B., 374 Hu, Zhengbing, 47, 281, 326 Huchenko, Inna, 208 I Iashvili, Giorgi, 117 Iavich, Maksim, 117, 127 Iosifov, I., 25 Iosifova, O., 25 Ivanishchev, Bohdan, 315 K Kalambet, Yaryna, 256 Kaplunov, Artem, 315 Kipchuk, F., 25 Korenko, Dmytro, 315 Korniyenko, Bogdan, 196 Kovalchuk, Pavlo, 89 Kovalchuk, Volodymyr, 89 Kravets, Petro, 304

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 Z. Hu et al. (Eds.): ICCSEEA 2021, LNDECT 83, pp. 467–468, 2021. https://doi.org/10.1007/978-3-030-80472-5

468 Krucheniuk, Anatolij, 15 Kvachadze, N., 127 L Ladieva, Lesya, 196 Lastovsky, Orest, 270 Lechachenko, Taras, 75 Lemeshko, Oleksandr, 47 Lendyuk, Taras, 138 Levkivskiy, Ruslan, 102 Li, Zhiwen, 175 Lipyanina, Hrystyna, 138 Lishchuk, Kateryna, 232 Litovchenko, Olena, 401 Liu, Yue, 444 Liu, Zhiping, 434, 444 Liubchuk, Olha, 363 Loutskii, Heorhii, 315 Lytvyn, Vasyl, 304 M Ma, Jie, 411 Maslov, Vadym, 290 Mastykash, Oleg, 56, 256 Matiash, Tetiana, 15 Mgebrishvili, N., 127 Miroshnychenko, Nelia, 401 Moiseev, Gr., 127 Moiseienko, Vladyslav, 3 Monashnenko, Anna, 147 Movchan, Kostiantyn, 219 Mukha, Iryna, 232, 389 N Nazaruk, Maria, 75 Nosov, Pavlo, 3 O Olefir, Oleksandr, 270 Oleshchenko, Liubov, 219 Osolinskiy, Oleksandr, 138 Ospanova, Dinara, 326 P Pashko, Anatolii, 37 Pavlov, Alexander, 232 Perova, Iryna, 401 Petrashenko, Andriy, 290 Petrushenko, Natalia, 363 Popovych, Ihor, 3 Prokopenko, Igor, 347

Author Index Q Qiu, Junan, 423 Qiu, Qizhi, 423 R Rao, Wenbi, 175 Rehida, Pavlo, 315 Romanovskyi, O., 25 Rozhko, Viktoria, 89 Rozora, Iryna, 37 Rybachuk, Liudmyla, 232 S Sachenko, Anatoliy, 304 Sachenko, Oleg, 304 Sachenko, Svitlana, 138 Shapovalova, Anastasiia, 47 Sharko, Marharyta, 363 Sherstjuk, Volodymyr, 102 Shilinh, Anna, 56 Simakhin, Volodymyr, 326 Sokol, Igor, 102 Sokolov, V., 25, 374 Sukaylo, I., 25 Sydorov, Nikolay, 244 Syerov, Yuriy, 56, 256 Syniavska, Olga, 37 T TajDini, M., 374 Tovstokoryi, Oleh, 3 Tyshchenko, Oleksii K., 281 V Volokyta, Artem, 315 Vysotska, Victoria, 304 W Wang, Yanan, 454 Y Yakymchuk, Tetiana, 363 Yatskiv, Vasyl, 138 Yeremenko, Oleksandra, 47, 401 Yevdokymenko, Maryna, 47 Z Zarichkovyi, Oleksandr, 389 Zhang, Yi, 175, 454 Zharikova, Maryna, 102 Zhou, Xiaofen, 454 Zhussupekov, Muratbek, 162 Zinchenko, Serhii, 3